Mexico Is Ageing on an Income It Cannot Retire On, and the Budget Already Shows It

Why a fertility rate of 1.2 and a pension bill above 6% of GDP measure the same problem, and the two indicators that could prove this wrong

AsiaPresswire – Mexico’s fertility rate was 1.9 children per woman in 2019. In 2024 it was 1.2. Replacement is 2.1. In five years the country moved from a fertility profile that roughly holds a population steady to one where each generation is 43% smaller than the one before it, and it did so at a income level where the state has not finished building the systems that older people need.

That is the whole argument, and it can be falsified in two places. Either fertility recovers toward replacement, or the formal labour market expands fast enough to fund the pension promises already written. If neither happens by 2028, the arithmetic below is not a forecast, it is a schedule.

The Number That Turned

The 2025 inter-census survey, reported by INEGI and relayed through Mexican statistical coverage in 2025-26, put the population at 130.4 million with 19.2 million people aged 60 or over. That is 14.7%. The median age is 32. The ageing index, which counts people aged 60 and over per 100 children aged 0 to 14, reached 68.2, against 47.7 in 2020. Life expectancy at birth is 75.5 years.

None of those numbers is unusual for a middle-income country in isolation. What is unusual is the speed and the sequencing. A fertility rate that falls 37% in five years is not a slow demographic drift; it is a behavioural break, and it arrives before the institutions that amortise it.

Table 1: Six demography indicators, and what each one implies

Indicator	Value	Year	What it changes
Total population	130.4 million	2025	Denominator for every per-capita claim
Population aged 60 and over	19.2 million	2025	14.7% of the total
Median age	32 years	2025	Labour supply still growing, briefly
Ageing index	68.2 per 100 children	2025	Was 47.7 in 2020
Fertility rate	1.2 children per woman	2024	Was 1.9 in 2019
Life expectancy at birth	75.5 years	2025	Duration of the payout

Read the median age and the fertility rate together and the trap appears. A median age of 32 means Mexico still has a large working-age cohort, which is why the fiscal problem does not feel urgent in 2026. A fertility rate of 1.2 means that cohort is the last large one. The two facts are consistent, and they point in opposite directions for about fifteen years.

Three Variables, Not One

Most commentary on Mexican ageing collapses into a single variable, usually the dependency ratio, and then argues about projections. The dependency ratio is a consequence. Three variables drive it, and only three matter for the next decade.

Fertility. At 1.2, Mexico is below every OECD member except Korea and Italy, and it got there faster than either. Fertility is the slowest variable to reverse and the fastest to move; it responds to housing costs, female labour participation and contraceptive access, none of which a pension reform touches.

The contributory and non-contributory split. Mexico runs two pension systems that behave in opposite ways. Contributory pensions, financed by worker and employer contributions through IMSS and ISSSTE, are contractually fixed and growing with the number of retirees. Non-contributory transfers, financed from the budget, are discretionary and politically expandable. In 2026 the budget froze the first and expanded the second, and that is the decision that matters.

Coverage. The share of the population covered by public health insurance fell from 77.1% in 2020 to 72.5% in 2025. Coverage is the variable that determines whether an ageing population shows up as fiscal cost or as household cost. When coverage falls while the old population grows, the cost does not disappear. It moves to the family, and then to the private clinic.

What the 2026 Budget Actually Did

The 2026 expenditure budget allocates 2.3 trillion pesos to pensions, an increase of 3.7% overall. That headline number hides the structure.

Contributory pensions, the ones that depend on contributions from workers, employers and the state, rise by 0.05%. The IMSS pension budget falls 2.5%. The ISSSTE budget falls 2.3%. The number of pensioners grows about 3% a year. Meanwhile the non-contributory old-age pension rises 5.2%, the disability pension 20.8%, and the new pension for women aged 60 to 64 rises 266%, expanding to three million beneficiaries.

Table 2: The 2026 pension budget, by line

Line	2026 change	Beneficiaries	Political status
Total pension spending	+3.7%	—	2.3 trillion pesos
Contributory pensions	+0.05%	about 3% more each year	Contractual
IMSS pension budget	-2.5%	largest block of retirees	Contractual
ISSSTE pension budget	-2.3%	public-sector retirees	Contractual
Old-age pension, 65 and over	+5.2%	13.7 million	Discretionary
Disability pension	+20.8%	2.0 million	Discretionary
Pension for women aged 60-64	+266%	2.8 million registered	Discretionary
Total social programme beneficiaries	—	42.8 million	—

Two things follow, and they are not ideological.

First, the contributory system is being asked to deliver a 3% larger caseload on a budget that grows by five hundredths of one percent. That is arithmetically a cut in the real value of each pension, before inflation. Divide 1.0005 by 1.03 and the real value per beneficiary falls 2.9%. Add 4% inflation and the fall is 6.6% in a single year.

Second, the non-contributory system is becoming the larger growth engine. Thirteen point seven million people receive 6,400 pesos every two months, which is 3,200 pesos a month, about US$171 at 18.7 pesos to the dollar. The women’s pension pays 3,100 pesos every two months, 1,550 a month, about US$83. Multiply by beneficiaries and the two transfers cost roughly 578 billion pesos a year, or 25% of the entire pension budget.

That 25% is the number to hold onto. A quarter of Mexico’s pension money now moves through programmes that a future government can change with a budget line, and three quarters moves through contracts that a future government cannot. The fiscal risk is not that pensions are too generous. It is that the rigid part is frozen while the flexible part expands, and the flexible part is the one that reaches the poorest old people.

The Arithmetic of a Frozen Line

Here is where an ageing population becomes a budget problem rather than a demographic one.

The pension bill passed 6% of GDP in the 2026 budget and approaches a quarter of net public spending, according to the Centre for Economic and Budgetary Research. Six percent of GDP is not remarkable against the OECD, whose public pension spending sits in the region of 8%. What is remarkable is the ratio to the old-age share.

Divide 6% by 14.7% and Mexico spends about 0.41 percentage points of GDP for every percentage point of population aged 60 and over. Do the same for an OECD average of roughly 8% of GDP and roughly 24% of population aged 65 and over, and the figure is 0.33. Mexico is spending about a quarter more per old person, relative to its output, than the rich-country average.

The comparison has an obvious flaw and it runs in one direction. Mexico’s 14.7% counts people aged 60 and over, while the OECD’s 24% counts 65 and over. Using the higher threshold for Mexico would shrink the denominator and raise the ratio further, so the flaw makes Mexico look cheaper than it is, not more expensive. The conclusion survives the correction.

Table 3: Pension spending against old-age share

Measure	Mexico, 2026	OECD reference	Ratio
Public pension spending, % of GDP	6.0	about 8.0	0.75
Population above the threshold, %	14.7 (60+)	about 24 (65+)	0.61
Spending per point of old-age share	0.41	0.33	1.23
Share of net public spending	about 25%	—	—

A country that spends like a rich country on pensions while collecting taxes like a middle-income one has three options, and it is worth naming them plainly because every Mexican budget since 2020 has chosen among them implicitly.

• Raise contributions or taxes, which hits the formal sector that is already paying.
• Cut the real value of contributory pensions, which is what a 0.05% increase against 3% more pensioners does by default.
• Expand non-contributory transfers and accept that the fiscal cost grows with the elderly population rather than with the wage bill.

The 2026 budget chose the second and the third at the same time. That is a coherent political strategy and an incoherent fiscal one, because the two choices have different growth rates.

The ratio has a clock on it, and the clock is easy to read. If the number of pensioners grows 3% a year while the number of registered workers grows 1%, the ratio of contributors to retirees falls 1.9% a year, and 17.8% over a decade. That is not a crisis scenario; it is the current trend extended. To hold the ratio flat, registered employment has to grow at the same 3% as the pensioner count, which in a country where informal work still accounts for a majority of employment is a labour-formalisation target rather than a growth target. Framed that way, the pension debate and the informality debate are one debate, and the 2026 budget treated them as two.

Who Is Hit First: Women Aged 60 to 64

The distributional answer is unusually clear, and it comes from the design of the new programme rather than from advocacy.

A woman who turns 60 in Mexico in 2026 and who spent her working life in the informal economy has no contributory pension, because she never had an employer contributing on her behalf. She becomes eligible for the women’s pension at 60, four years before the universal old-age pension at 65, and she receives 1,550 pesos a month, about US$83.

Two point eight million women are registered. The budget for that line rose 266% in one year, which is the largest percentage increase in the entire pension bill and the clearest statement of political priority in the 2026 budget. It is also a recognition of a real gap: Mexican women have lower formal labour participation, more interrupted contribution histories, and a longer life expectancy than men, which means they draw a smaller pension for more years.

The programme is means-tested by age and gender, not by contribution history, so it reaches exactly the group the contributory system never covered. That is its strength. Its weakness is that 1,550 pesos a month is not a pension. It is a transfer that covers part of a household’s food bill, and it will be paid for as long as the budget line exists.

The second cohort is less visible. Workers hired before July 1997 fall under the 1973 pension law, which calculates the benefit from the average salary of the last five years rather than from an individual account. That group had a legal route to raise its own pension by making voluntary contributions at a higher declared salary, known as Modalidad 40.

On 22 October 2026 that route narrowed. New entrants can only contribute on the last salary registered with the institute plus one unit of measurement, with the UMA set at 117.31 pesos a day for 2026. Retroactive payments from before the application date are no longer possible, and self-employed workers registered under Modalidad 10 cannot also take Modalidad 40. For a worker who was planning to spend five years raising his declared salary before retiring, the option closed on a specific date, and the difference between the pension he expected and the one he will receive is not a projection. It is a subtraction.

Two Indicators That Are Actually Independent

Forecasts about ageing are cheap. What is expensive is knowing which number to watch, and the two numbers below are chosen because they move for different reasons.

The first is the ratio of IMSS-registered workers to the population aged 60 and over. This is the contributory base divided by the people it must support, and it is published monthly. If the ratio rises, formal employment is outrunning ageing and the contributory system can hold. If it falls, every promise in the 1973 and 1997 laws becomes a claim on general revenue.

The second is private spending on geriatric care. IMARC put Mexico’s geriatric healthcare market at US$20.58 billion in 2025. This indicator measures the cost that has already been pushed out of the public system into households. It is independent of the first because it responds to coverage, not to employment: when public health coverage falls from 77.1% to 72.5%, the private market grows whether or not more people have formal jobs.

Table 4: The two indicators, and what each one is measuring

Indicator	Source	Frequency	What a move means
IMSS-registered workers per person aged 60+	IMSS	Monthly	Contributory solvency
Private geriatric care spending	IMARC and clinic filings	Annual	Cost shifted to households
Public health coverage share	INEGI	Annual	Whether the shift accelerates
Ageing index	INEGI	Annual	Speed of the demographic turn

Watch the first two together. If registered workers per old person is stable and private geriatric spending is growing at double digits, Mexico is not facing a pension crisis; it is facing a care crisis, and the money is already leaving households to pay for it. If both deteriorate, the pension problem and the care problem arrive in the same fiscal year, which is the scenario that the 2026 budget’s structure makes likely.

What Would Prove This Wrong

Three observations would break the argument, and each is measurable within two years.

• Fertility recovers above 1.6. A rebound of that size would require either a large-scale childcare expansion or a change in housing costs that no current policy implies. If it happens, the 2045 dependency ratio is materially better than the schedule above.
• Formal employment grows faster than the old population for three consecutive years. IMSS registrations have grown in most years since 2010, but rarely faster than 3% net. Three years above 5% would mean the contributory base is winning.
• Remittances to households headed by someone over 65 rise sharply. Mexico received record remittances through the 2020s. If those flows are increasingly directed at elderly households, the household sector is absorbing the coverage gap and the fiscal crisis is deferred, not solved.

Put a number on the second observation. Registered employment growing 5% a year against a pensioner count growing 3% raises the contributor-to-retiree ratio 19.4% over ten years, which is enough to carry the 1973 cohort through its peak drawdown without a real cut. That is the size of the move required, and it is why the formalisation rate rather than the fertility rate is the variable to argue about in the next two budgets.

Two of the three are about the labour market, which is the variable that Mexican pension debates consistently underweight. The demographic schedule is set. What is not set is how many people are paying into the system while the schedule unfolds.

Data Basis, Two Limits, and the Number I Could Not Find

The demographic figures come from INEGI’s 2025 inter-census survey as reported in Mexican statistical coverage, and the ageing index comparison uses the 2020 census as the base.

• Population 130,392,389, aged 60 and over 19,204,670, median age 32, ageing index 68.2, life expectancy 75.5 and public health coverage 72.5% are 2025 INEGI figures, with the 2020 comparators of 47.7 for the ageing index and 77.1% for coverage stated in the same releases.
• Fertility of 1.2 in 2024 against 1.9 in 2019 is INEGI data reported through Mexico News Daily.
• The 2.3 trillion peso pension total, the 3.7% overall increase, the 0.05% contributory increase, the 2.5% and 2.3% cuts to IMSS and ISSSTE, and the 6% of GDP and quarter-of-net-spending ratios are from the Centre for Economic and Budgetary Research’s analysis of the 2026 expenditure budget.
• The 13.7 million old-age beneficiaries at 6,400 pesos bimonthly, the 2.8 million women registered at 3,100 pesos bimonthly, the 266% increase and the 42.8 million total social programme beneficiaries are from the Welfare Ministry’s 2026 briefing.
• The Modalidad 40 changes, the one-UMA limit and the 117.31 peso daily UMA for 2026 are from the IMSS decree of 21 October 2026 as reported by El Pais on 22 October 2026.
• The US$20.58 billion geriatric healthcare market is IMARC Group’s 2025 estimate.

Two limits deserve stating. The first is that INEGI’s 2025 figures come from an inter-census survey rather than a census, so the denominators carry sampling error that the published tables do not fully express; the ageing index in particular is sensitive to the estimate of the 0-14 population, which is the hardest group to count. The second is that the OECD comparator uses a 65-and-over threshold against Mexico’s 60-and-over, which I have flagged rather than corrected, because correcting it properly requires age-specific spending data that neither the OECD nor the Mexican finance ministry publishes at that granularity.

One number I could not find is the one that would settle the care question: the share of Mexican households with at least one member over 65 that report out-of-pocket health spending above 30% of household income. The national health and nutrition survey collects the components but not in a form that is published annually at that threshold. Until it is, the private geriatric market size is the best available proxy, and a market size is a weaker instrument than a household survey.

Sources

• INEGI, population and housing statistics — 2025 inter-census population, ageing index and median age
• INEGI data via Mexico News Daily — fertility rate of 1.2 in 2024 and coverage trends
• INEGI data via Puebla Online — population 130.4 million and 19.2 million aged 60 and over
• Centre for Economic and Budgetary Research, analysis of the 2026 budget — pension spending above 6% of GDP and the IMSS and ISSSTE allocations
• Pension Policy International, pensions, retirement age and gender inequality — PEF 2026 pension share of GDP and of net spending, OECD comparison
• Welfare Ministry briefing via Mohuawang — 13.7 million old-age beneficiaries, 2.8 million women registered, 42.8 million total
• El Pais, changes to IMSS Modalidad 40 — decree of 21 October 2026, one-UMA limit and UMA value
• IMARC Group, Mexico geriatric healthcare market — US$20.58 billion market size for 2025

This article was prepared by the expert team at Global Infinity Labs, drawing on insights and data provided through its MCP data services.

About Global Infinity Labs

Global Infinity Labs is an AI innovation company specializing in MCP data services.

Email: support@globalinfinitylabs.com

source https://newsroom.seaprwire.com/press-releases/finance/mexico-is-ageing-on-an-income-it-cannot-retire-on-and-the-budget-already-shows-it/

Satlyt’s In-Orbit Scheduling Test Worked. Sharing Compute Between Satellites Is Still Unbuilt

SeaPRwire – The experiment Satlyt ran earlier this month used exactly one computer in space. It sat aboard a spacecraft, worked through a simulated network of four computing hosts, four users and eight services, and sent its answers back to Earth. The other three machines in the scenario existed only as numbers on that one.

What the run establishes is narrower than the phrase “space data center” implies, and more useful than a demo. The software produced a scheduling plan on orbit, using the compute that was actually available where the hardware physically was, instead of modelling the problem on a ground machine and calling that a result. Thirteen months of engineering went into reaching that point.

The scheduling model works through declarations rather than approvals. Operators set permissions, priorities and operating limits, and the software assigns tasks inside those bounds. Afullo’s description of it is precise on this point: users “will only need to define permissions, priorities, and operating limits.” Nobody signs off on each assignment. The operator writes the fence, and the system stays inside it.

That design choice carries the main unresolved question for this class of software. A fence wide enough to be useful is also wide enough to permit an assignment the operator would have refused if asked. The test did not probe that boundary, because probing it requires workloads and hardware that the single-computer setup did not have.

The result Afullo reports is a trade-off instead of a win. One of the approaches available to the scheduler performed best and consumed more computing power than the alternatives. On a spacecraft, where power and thermal headroom are the binding constraints and neither can be upgraded after launch, “best” and “affordable” rank differently. A scheduler that ignores the difference produces plans the satellite cannot execute.

The test also turned up a defect of the kind that a ground simulation tends to hide. The software was stamping communication delay by arrival time, the moment data showed up, rather than by elapsed delay, how long it took to get there. The two numbers diverge exactly when a scheduler has to decide between waiting for a link and reassigning the work to another computer. Afullo treats the fix as a correction, not a failure, and the record supports him: development, onboard execution and the return of outputs all completed.

Context for how much of the problem this solves: NASA already operates a network built to move data through space when links are delayed or intermittent. It went into use earlier this year and has been tested on aircraft. That layer answers whether bits arrive. Satlyt’s layer answers who computes what, and its answer has to change as the link changes. The second question is only interesting once the first has a working answer, which is roughly where satellite operators are now.

The wildfire example is the one Afullo reaches for, and it is a fair illustration of the economics rather than a description of a deployed product. A spacecraft that sees a fire can send raw imagery down and wait for a ground team, spending a communication window on pixels, or it can run detection onboard and send a shorter answer. Earth observation operators have been doing onboard processing for years. What Satlyt is selling is not detection; it is the coordination layer that would let one job be split across several spacecraft when no single one has the compute or the connectivity to finish it alone.

Its software is on three satellites in orbit. The research behind the test was NASA-funded and run with the University of Houston, and the agency has not selected Satlyt for a mission or chosen to deploy the technology. That distinction matters for reading the result. This is a company validating its own planning component inside an agency-funded research project, not an operational system whose output anyone depends on yet.

The list of what remains, in Afullo’s own account, is specific: usable communication links, permission from the operators of participating spacecraft, secure transfers between them, resource scheduling across a fleet, and recovery when a node drops out. The first item on that list is not a software problem, and the second is not either. Cross-operator compute sharing asks satellite owners to accept that another company’s scheduler will direct work on their hardware, under limits they wrote but cannot watch in real time.

Multi-spacecraft transfer is the untested part. Every result so far comes from a single physical machine reasoning about machines that were not there. Satlyt says it will test its software under various workloads, hardware and connectivity conditions before attempting that, and it is researching applications beyond low Earth orbit, including 4K video streaming from the Moon through the NASA-funded work.

The company raised $8 million and reached the top five of TechCrunch Disrupt Battlefield this year, which is a reasonable amount for a scheduling layer and a small amount for a constellation. The longer-term ambition Afullo describes is the standard platform argument: developers write applications for orbit without rebuilding infrastructure for each mission. It is also the hardest version of the argument to deliver in a domain where the hardware is fixed at launch and the network topology changes with orbital mechanics.

Hardware heterogeneity is the constraint that has not been priced. A scheduler assigning work across spacecraft has to model what each node can do, and satellites launched in different years carry different processors, different power budgets and different thermal designs. The test avoided this by simulating four hosts that were interchangeable. Operators running a mixed fleet, which is most of them, would be asking the software to plan across machines it has never benchmarked, and the plan is only as good as the assumptions behind it.

The next milestone is a demonstration that spans separate spacecraft. Until one exists, the accurate description of Satlyt’s software is a scheduler that has been shown to generate a plan on one computer in orbit, and has not yet been shown to work as a network.

source https://newsroom.seaprwire.com/press-releases/technologies/satlyts-in-orbit-scheduling-test-worked-sharing-compute-between-satellites-is-still-unbuilt/

What the Ofcom Information-Notice Ruling Actually Decides, and What It Leaves Open

SeaPRwire – Five grounds went into the Divisional Court. Four of them never reached a hearing on their merits. That procedural detail is the part of the 30 October judgment worth understanding first, because it decides how much of the platforms’ argument was actually tested and how much was simply refused at the gate.

Under the Online Safety Act 2023, section 100 lets Ofcom require regulated services to hand over information the regulator needs to carry out its functions. Meta and WhatsApp, X and TikTok each received notices and each brought judicial review. The judgment, from Lord Justice Singh and Mrs Justice Collins Rice, is reported as [2026] EWHC 2779 (Admin), and it is the first judicial examination of that power.

The four grounds that failed at the permission stage were the argument that the notices circumvented the Act’s separate transparency regime, the argument that the information was not genuinely “required,” the argument that Ofcom had skipped an impact assessment under section 7 of the Communications Act 2003, and, for TikTok alone, predetermination. Permission was refused because the court considered those grounds unarguable, which is a lower bar than being wrong: a ground refused permission does not get argued out in open court, and the reasoning behind the refusal is correspondingly thin.

Proportionality survived that filter. The companies framed it under both section 100(4) and the European Convention on Human Rights, the court granted permission, heard it in full, and rejected it. So of five complaints, exactly one was examined on the evidence, and it lost.

The finding inside that single ground is where the judgment will be cited. The platforms said compliance would impose an extraordinary burden. The court found the evidence offered in support lacked “pertinent substance,” meaning it did not show in concrete terms why the demands were excessive. Read that as a statement about the quality of the record rather than about the size of the burden: the court did not conclude the burden was small, it concluded the platforms had not demonstrated it.

For anyone advising a regulated service on how to resist a section 100 notice, that is the operative paragraph. Expense and complexity are not self-proving. A notice running to hundreds of pages may well cost a great deal to answer; a court will want the costed version, itemised against the specific requests, before it weighs that cost against the regulator’s need. The judgment rewards a company that can produce engineering time estimates, headcount allocations and per-request breakdowns, and penalises one that asserts scale.

What the ruling does not establish is more extensive than what it does. It says nothing about whether the information Ofcom sought was the information it should have sought, only that its decision on that question survives a proportionality review on the record before the court. It says nothing about how Ofcom uses the material, or what standard applies when the data becomes the basis of a breach investigation. Those questions remain where they were.

The transparency argument deserves a note, because it is the one with the widest potential reach. The Act gives Ofcom a separate transparency regime with its own procedures for public reporting, and the companies argued that the information notices were an attempt to route around it. The court found that unarguable. If that conclusion holds through any appeal, the practical effect is that Ofcom can gather material for its own purposes without triggering the disclosure obligations attached to the transparency route, and the two channels stay legally distinct.

Two disputes sit outside this judgment. Meta’s separate challenge to the way fees and penalties are calculated by reference to worldwide revenue is still before the courts, and it addresses a different power: not what Ofcom may ask for, but what it may charge. The platforms may also seek permission to appeal the 30 October decision, which would put the proportionality finding in front of the Court of Appeal and test whether the evidential threshold was set correctly.

The wider read-across is a matter of institutional design rather than of this case. Section 100 is a demand power exercised without prior judicial authorisation. A regulator decides what it needs, issues a notice, and the company either complies or litigates after the fact, which is why the permission stage matters so much: the four grounds that never got heard are the four arguments a regulated service would most want to make. Jurisdictions drafting equivalent regimes will look at how little of the challenge survived, and at how narrow the surviving ground turned out to be.

One asymmetry is worth naming. Because four grounds were refused permission, the court’s reasoning on them exists mainly as a refusal rather than as a reasoned judgment, and a refusal is harder to appeal and harder to rely on in the next case. The proportionality finding rests on a full hearing and a developed record, which makes it the part later litigants can actually use. That is an odd result for a first test of a statutory power: the narrowest ground produced the most durable holding.

What to watch is unglamorous. Whether an appeal is filed. Whether the pending fee challenge produces a different result on a power the court has not yet construed. Whether the data collected under these notices shows up in enforcement action, which is the only test of whether the information was worth demanding. And whether the next company served with a section 100 notice answers it, or spends the effort to build the costed record that this judgment says is required to resist one.

The notice power is not self-executing either. What a platform must produce, in what format and by what date are set by the notice itself, and nothing in this judgment constrains how broadly the next one may be drawn. A ruling that upholds one set of demands says little about the scope of the following set.

The ruling tells regulators their information powers will be enforced and tells companies what kind of evidence it takes to push back. It does not tell anyone whether the requests were well judged, and the court was not asked to answer that.

source https://newsroom.seaprwire.com/press-releases/policy-analysis/what-the-ofcom-information-notice-ruling-actually-decides-and-what-it-leaves-open/

Mistral Released the Weights and Handed the Safety Burden to Whoever Runs Them

SeaPRwire – A model release is usually a product event. This one is a procurement event. Mistral put the weights for Mistral Large 4, a trillion-parameter model it calls Le Chonk, on Hugging Face and behind its own API on Thursday, which means a customer can now take the checkpoint, install it on hardware it controls, and stop asking permission to use it. The company’s pitch is that this matters most in cybersecurity, where the buyer is a security team that wants to point a model at its own code and its own logs.

Weights are the numerical result of training. Publishing them does not publish the data, the recipe, or the evaluation harness, and it does not oblige Mistral to keep shipping updates. What it does is remove a switch. An API customer’s access depends on a vendor’s continued willingness to serve that customer, at that price, under that policy. A weights customer’s access depends on hardware and electricity. Stock, Mistral’s vice president of science, put the claim in almost exactly those terms: access “can’t be cut off by definition because the model is open-weighted.”

That sentence is the whole strategy, and it is worth being precise about what it guarantees. It guarantees the checkpoint. It does not guarantee a patched checkpoint, a model that improves, or a vendor who answers when something goes wrong at 3 a.m. during an incident. Organisations that adopt open weights for continuity are trading a vendor dependency for an internal maintenance obligation, and the second one is easier to control but harder to staff.

The cybersecurity framing is where the argument gets strongest and least examined. Stock’s case is that closed models ration the work defenders need to do: they rate-limit, and they refuse requests that look like offensive security even when the requester is trying to find a bug in its own systems. A model running on-premises has no provider in the loop to refuse anything. The same property that makes it useful to a defender makes it useful to anyone else, which is the part of the release Mistral addresses with a claim about who benefits first.

“Some malicious actors might try to actually leverage their own technology to attack,” Stock said, arguing that defenders therefore need the capability “as soon as possible.” That is a race argument, not a safety argument, and it has a testable implication: if the defenders who receive the weights are not actually faster than the attackers who can download the same file, the release has moved capability around without changing the balance. Nobody in the announcement offers a way to measure which happened.

Reversal is the other asymmetry. An API can be withdrawn in an afternoon; a checkpoint that has been mirrored cannot be. If a later revision of the model turns out to behave badly on offensive tasks, Mistral can stop shipping it and cannot stop anyone from running the copy already on disk. That is the price of the guarantee Stock is selling, and it is paid by everyone who shares a network with the buyer, not only by the buyer.

The performance claims need the same treatment. Mistral says the model is comparable to the strongest rival open-weight release of recent months and uses 47 percent fewer tokens per task. Token efficiency is not a vanity metric for on-premises buyers; it is the difference between a cluster that fits in a budget and one that does not. But the comparison is Mistral’s own, published with the release, and the 47 percent is an average across whatever task set the company chose.

Separately, Mistral claims the model outranks every other open model on the Coding Agent Index maintained by Artificial Analysis. One index, one vendor-supplied submission, no independent replication in the announcement. It is a real signal — third-party leaderboards are hard to game completely — and it is also a single number standing in for a capability that buyers will experience as a distribution of failures across many repositories.

Mensch’s framing is philosophical: AI should be “a tool for empowerment, not dominance,” and the model reaches “state-of-the-art performance” at “a fraction of the compute and capital committed by closed labs.” The compute claim is the one that would matter most to an investor and the one least verifiable from outside, because nothing in the release states the training budget in a form anyone can check. What can be checked is the download.

There is a quieter commercial logic underneath. A lab that cannot outspend its rivals on training runs can still compete on distribution: if the default self-hosted model for regulated buyers is yours, you sell fine-tuning, support, and the paid tier for everyone who would rather not run a GPU cluster. Mistral’s own statement points at the buyer segment — enterprises and public institutions that want to “download, deploy, and own” the system — rather than at developers who will simply call an API.

That segment has its own constraints. A bank or a ministry deploying a trillion-parameter model on-premises needs the safety behaviour to be somebody’s job. With a closed API, moderation and abuse monitoring sit with the provider by default. With weights, they sit with the operator, who may not have a red team detached to the model and who now owns every refusal decision the model does or does not make. The release shifts capability to the buyer and the responsibility with it, and Mistral’s announcement does not describe what it hands over to help with the second half.

Stock was asked about rogue AI agents and declined to escalate. “We don’t want to fearmonger people. We think it’s a problem that is solvable. It’s a hard problem to have good security.” A lab releasing cyber-capable weights has an obvious interest in that position, and the position is not wrong on its own terms; it is simply not evidence. The evidence would be an evaluation of what the released checkpoint actually does when someone points it at a target without authorization, run by people who do not work for Mistral.

Two things would settle more than the press release does. First, independent red-team results on the released weights, published with methodology, including the refusal rate on offensive tasks rather than a summary adjective. Second, a public account of how many regulated buyers run it in production a year from now and what broke. Until then, the honest description of this release is narrow: a competent lab has made a frontier-scale checkpoint available to anyone who can host it, and has bet that the defenders will be the ones who bother.

This article was prepared by the expert team at Global Infinity Labs, drawing on insights and data provided through its MCP data services.

About Global Infinity Labs

Global Infinity  Labs is an AI innovation company specializing in MCP data services.

Email: support@globalinfinitylabs.com

source https://newsroom.seaprwire.com/press-releases/technologies/mistral-released-the-weights-and-handed-the-safety-burden-to-whoever-runs-them/

Saudi Payments After 85%: The Fight Moves to the Terminal

SAMA data show electronic payments at 85% of retail while POS value per terminal falls 9.4% in a quarter; the constraint is now unit economics.

AsiaPresswire – Saudi Arabia won the argument it spent a decade having. Electronic payments reached 85% of retail transactions in 2025, against a Financial Sector Development Program target of 70% for that year, and against 57% in 2021. The number is now quoted in every fintech deck in Riyadh, usually on slide two.

The number also marks the end of a growth story and the start of a different one. Once four out of five retail transactions are electronic, the remaining 15% is not a market waiting to be converted; it is a residue of small vendors, cash-preferred customers and transactions that are cheaper to settle in notes. Growth from here has to come from somewhere else.

The place it has to come from is visible in the same statistical release, and it is less comfortable than the headline. Saudi POS sales fell from SAR 189.7 billion in the first quarter of 2026 to SAR 179.0 billion in the second, while the number of POS terminals rose from 2.4 million to 2.5 million. Divide one by the other and value per terminal drops 9.4% in a single quarter. That is the number to plan around.

What the 85% actually closed

Start with what the adoption curve was measuring. Between 2021 and 2025 the electronic share of retail transactions went 57%, 62%, 70%, 79%, 85%. Each of those steps was produced by a specific intervention: mandated acceptance for government payments, wage protection rules that pushed payroll into bank accounts, a national QR standard, and the mada debit card as the default instrument.

That work is done. The 15% that remains in 2026 is the hard tail, and it is hard for a reason that has nothing to do with technology. A bakery selling SAR 4 loaves pays a percentage fee on each card transaction, absorbs the settlement delay and keeps a paper trail. Cash costs nothing per transaction and leaves no trail. No amount of terminal subsidy changes that arithmetic until the fee per transaction falls below the value of the anonymity.

So the useful question for 2026 is not how to convert the last 15%. It is what happens to the economics of the 85% now that growth has to come from volume and pricing rather than from new users.

Three variables that now set the ceiling

Three quantities in the SAMA quarterly release determine whether the payments business in Saudi Arabia gets easier or harder over the next eight quarters. None of them is the electronic share.

The first is value per terminal. It measures how much money flows through each acceptance point per quarter, which is the base on which an ad valorem merchant fee is charged. If it falls, every acquirer’s revenue per installed device falls with it unless the rate rises, and the rate is under competitive pressure from new entrants.

The second is cash withdrawn per ATM transaction. It measures whether cash is being used for smaller, more frequent purposes or for larger, less frequent ones. The direction matters more than the level, because it determines refill frequency and therefore the cost per riyal moved.

The third is cards per resident. It measures whether issuance still has room to grow. Once a market has roughly two cards per person, new cards stop adding spend and start adding cost.

Table 1: SAMA payment indicators, first and second quarter of 2026, with unit ratios calculated here

Indicator	Q1 2026	Q2 2026	Change
POS transactions (million)	2,982.6	3,153.7	+5.7%
POS sales (SAR billion)	189.7	179.0	-5.6%
POS terminals (million)	2.4	2.5	+4.2%
POS sales per terminal (SAR, quarter)	79,042	71,600	-9.4%
ATM transactions (million)	359.0	401.6	+11.9%
ATM cash withdrawn (SAR billion)	136.8	125.4	-8.3%
Cash per ATM transaction (SAR)	381	312	-18.1%
Cards issued (million)	66.9	69.2	+3.4%

 

Source: SAMA , Key Economic Developments Q1 and Q2 2026; ratios calculated from the published totals

Divide the totals by the terminals

The first variable is the one that decides whether an acquiring business is worth owning.

In the first quarter of 2026, Saudi Arabia’s 2.98 billion POS transactions carried SAR 189.7 billion of sales across 2.4 million terminals. That is SAR 79,042 per terminal for the quarter, or about SAR 26,300 a month. In the second quarter, 3.15 billion transactions carried SAR 179.0 billion across 2.5 million terminals, or SAR 71,600 per terminal for the quarter.

The transaction count per terminal barely moved: 1,243 in the first quarter, 1,261 in the second, a gain of 1.5%. So this is not a story about terminals sitting idle. It is a story about the average basket getting smaller. Average POS ticket fell from SAR 63.60 to SAR 56.76 between the two quarters, a drop of 10.8%.

Two readings are possible, and they have opposite implications.

Reading one: merchants are pushing small purchases onto cards that used to be cash. A SAR 12 coffee paid by card instead of a note lowers the average ticket while raising volume, which is the intended outcome of the whole national programme. Under this reading, falling ticket size is a success metric.

Reading two: the marginal terminal being installed in 2026 is at a small merchant with low throughput, and the existing base is competing on price. Under this reading, falling ticket size is dilution.

The data cannot separate the two yet. What it can do is tell an operator which way to hedge. If reading one is right, revenue per terminal stabilises as the newly terminalised merchants mature. If reading two is right, revenue per terminal keeps falling and the installed base becomes a cost problem.

The cash side is not dying, it is shrinking in ticket size

The second variable is more interesting than the first because it contradicts the story everyone tells.

Cash withdrawals from ATMs fell 8.3% in value between the first and second quarters of 2026, from SAR 136.8 billion to SAR 125.4 billion. Over the same period the number of ATM transactions rose 11.9%, from 359.0 million to 401.6 million. The average withdrawal fell from SAR 381 to SAR 312, a drop of 18.1%.

More trips to the ATM, for less money each time. That is not the behaviour of a country abandoning cash. It is the behaviour of a country where cash has been demoted to a convenience: the taxi, the tip, the market stall, the school collection. Each of those needs a small amount of notes, and the notes have to be obtained somewhere.

For the banks that run the 14,400 ATMs reported in the second quarter of 2026, this is a cost problem dressed as a volume problem. The cost of keeping an ATM stocked is driven by the number of refill events, not by the value dispensed. If customers withdraw SAR 312 instead of SAR 381, the machine runs dry sooner for the same total value, and the cash-in-transit contract is priced per trip.

There is a straightforward calculation every ATM operator in the Kingdom should have run by now. Take the cost of one refill in 2026, divide by the value dispensed between refills, and you have the cost per SAR 1,000 dispensed. When the average withdrawal falls 18% and the number of withdrawals rises 12%, that cost per SAR 1,000 goes up by roughly a quarter, unless the refill schedule changes.

The third variable is already saturated

Cards issued reached 69.2 million in the second quarter of 2026, up from 66.9 million in the first. Against a resident population of roughly 35 million, that is about two cards per person, which is saturation for a market where most adults hold a mada debit card and a large share hold at least one credit card.

Issuance growth of 3.4% in a quarter is not nothing, but it is now adding cards to people who already have one. The incremental card does not create incremental spend; it fragments the same wallet across more instruments. For issuers, that means the cost of acquisition keeps rising while the revenue per card stays flat.

This is the point where a payments market stops being an adoption story and becomes a share story. Every riyal of growth has to be taken from another institution, another rail or another fee line.

The merchant side of the same arithmetic

A restaurant processing 1,000 card tickets a quarter at the current average of SAR 56.76 turns over SAR 56,760. At an ad valorem fee anywhere between 0.5% and 2%, that is SAR 284 to SAR 1,135 for the quarter, against a terminal rental and connectivity charge that is largely fixed. The smaller the ticket, the larger the share of the fee that fixed costs consume, which is why small merchants remain the strongest constituency for cash in 2026.

Run the same arithmetic against the average ticket of two years earlier and the fixed-cost share falls. The compression between the first and second quarters of 2026 therefore does two things at once: it raises the number of transactions a merchant must process to cover the same fixed cost, and it lowers the revenue each transaction generates for the acquirer. Both parties work harder for the same riyal.

The 371 firms chasing the same terminal

SAMA’s governor put the number of fintech companies operating in the Kingdom at 371 in September 2026, with 307 entities under the open banking framework and more than SAR 30 billion invested in the sector by the end of the first half of 2026. Eight new licences were granted in the second quarter alone, spanning BNPL, payment services, aggregation and open banking.

None of that capacity expands the terminal base much. It divides the same fee pool among more participants, which is how a maturing market transfers value from intermediaries to merchants. The 85% was built by adding acceptance points between 2021 and 2025. The next phase is built by removing intermediaries per transaction.

Where the growth actually went

One more division is worth doing, because it explains why the retail numbers feel flat while the system numbers look strong.

SARIE, the national interbank system, processed SAR 17,777.1 billion in the first quarter of 2026, up 16.1% year on year, and SAR 18,399.2 billion in the second, up 12.7%. Within that, customer payments were SAR 4,496.6 billion in the first quarter and SAR 4,333.7 billion in the second, with year-on-year growth of 8.2% then 2.2%. Interbank payments grew 20.5% then 18.9%.

Read those three lines together. Total value grows at a double-digit rate. Interbank value grows faster. Customer payments, the part that touches households and businesses, decelerated from 8.2% to 2.2% in one quarter.

The wholesale plumbing is where the growth is. That is consistent with a retail market that has finished converting and an economy that is still building: project finance, corporate treasury, government settlements and the capital flows behind the Vision 2030 pipeline all move through the interbank leg, not through a POS terminal.

For anyone whose business model depends on consumer transaction fees, this is the structural message of 2026. The consumer rail is mature, the wholesale rail is expanding, and the two have different customers, different pricing and different competitors.

Who gets hit first

Rank the participants by how quickly a 9.4% fall in value per terminal reaches their income statement.

• Acquirers and terminal deployers are first. Their revenue is a percentage of value processed, so a 9.4% fall in value per device is a 9.4% fall in revenue per device before any pricing response, partly offset by 4.2% more devices.
• Cash-in-transit and ATM operators are second. Their cost is driven by refill frequency, which rises when average withdrawals fall, so their margin compresses even as transaction volumes grow.
• Card issuers are third. Interchange is also ad valorem, but issuers earn on total volume, which is still rising, and they carry less terminal capital.
• Merchants are the beneficiaries, slowly. A market where acquirers are competing for shrinking value per terminal is a market where merchant discount rates have room to fall, and the merchants with enough volume to negotiate are the ones who will capture it.

SAMA’s licensing record points the same way. In the second quarter of 2026 the central bank granted licences to eight companies across BNPL, fintech consumer microfinance, payment services, finance aggregation and open banking. More licensees competing for the same terminal base is a recipe for fee compression, which is good for merchants and bad for anyone who built a model on stable take rates.

Two indicators, and what would prove this wrong

An argument about structure is only useful if it can be falsified. This one rests on two indicators that must be independent of each other, and they are: one measures the merchant side, the other measures the cash side. Neither is derived from the other.

The merchant-side indicator is quarterly POS sales per terminal, currently SAR 71,600. The claim in this piece is that it stays below SAR 78,000 through 2027 as small merchants join and ticket size compresses. The cash-side indicator is cash per ATM transaction, currently SAR 312. The claim is that it stays below SAR 350 over the same period, because cash has been demoted to small-ticket convenience.

Three observations would break the argument, and each is observable in the quarterly release.

First, if POS sales per terminal recovers above SAR 78,000 for two consecutive quarters, the compression was a seasonal artefact of the second quarter, not a structural trend. Second, if cash per ATM transaction climbs back above SAR 380, households are using cash for larger purchases again and the demotion story is wrong. Third, if customer payments through SARIE reaccelerate above 10% year on year while POS value per terminal keeps falling, the retail and wholesale legs are diverging for a reason this framework does not capture, most likely a shift of business spending out of cards and into transfers.

None of these is a small correction. Each one would mean the market is doing something other than maturing.

What to do on Monday

The analysis is only worth the decisions it changes. Three actions follow directly from the numbers above, and all three can be started this week.

• If you operate terminals, calculate your break-even ticket. Take your monthly cost per terminal, divide by your effective ad valorem rate, and you have the sales volume below which the device loses money. With average tickets at SAR 56.76, a meaningful share of your estate is probably below that line, and the fix is a fixed per-transaction floor rather than a higher percentage.
• If you run ATMs or cash logistics, re-price the refill contract against trips, not value. An 18.1% fall in average withdrawal with 11.9% more withdrawals means the trip count is the cost driver, and a contract priced on value dispensed will under-recover.
• If you are a merchant with more than SAR 5 million of annual card volume, ask three acquirers for a quote this month. Value per terminal is falling, eight new licensees were admitted in one quarter, and the pricing power has moved to your side of the table.

The 85% is a genuine achievement and it is also the end of an easy decade. What replaces it is arithmetic: fewer riyals per terminal, more trips to the ATM, and a wholesale rail that is growing faster than anything a consumer sees.

Data basis

Data as of October 2026. All levels come from SAMA’s quarterly Key Economic Developments releases for the first and second quarters of 2026, which report POS transactions and sales, ATM transactions and cash withdrawals, terminal counts, cards issued and SARIE values. The electronic share series of 57%, 62%, 70%, 79% and 85% for 2021 to 2025 is SAMA’s own, restated in the second quarter release. The 2025 electronic transaction count of 14.6 billion, against 12.6 billion in 2024, is also SAMA’s.

Five ratios are calculated from the published figures rather than quoted: POS sales per terminal, average POS ticket, cash per ATM transaction, transactions per terminal and cards per resident. Each divides one published quarterly figure by another from the same release, so the numerator and denominator share a period and a definition. Cards per resident uses an assumed resident population of 35 million, which is an approximation and not a figure from the payments release; a population of 33 million would give 2.1 cards per person and 37 million would give 1.9.

No seasonal adjustment is applied. The comparison is quarter against quarter within a single calendar year, which means it contains both the structural trend and any Ramadan and holiday effects that fall differently in Q1 and Q2.

Limitations

• Two quarters are not a trend. The Q1 to Q2 comparison spans Ramadan timing, which moves retail values independently of structure.
• Cards per resident assumes a population figure that SAMA does not publish in this release.
• Interbank and customer payment splits are reported gross, so they cannot be netted against retail turnover.

Sources

• SAMA , Key Economic Developments Second Quarter 2026
• SAMA , Key Economic Developments First Quarter 2026
• Argaam via Ajel , weekly POS report and the 2021 to 2025 electronic share series, August 2026
• Saudi Gazette via KSA News Today , March 2026 consumer spending, POS and ATM detail, April 2026
• Saudi Online , Money20/20 Middle East, SAMA governor on 85% electronic share and 371 fintechs, 15 September 2026

This article was prepared by the expert team at Global Infinity Labs, drawing on insights and data provided through its MCP data services.

About Global Infinity Labs

Global Infinity  Labs is an AI innovation company specializing in MCP data services.

Email: support@globalinfinitylabs.com

source https://newsroom.seaprwire.com/press-releases/finance/saudi-payments-after-85-the-fight-moves-to-the-terminal/

China’s Factories Are Not Leaving. The Last Three Steps Are.

A costed comparison of transshipment, overseas plants and equipment exports for Chinese manufacturers in 2026, and where each route fails.

SeaPRwire – In March 2026, US Customs and Border Protection opened an enforcement case over Chinese golf carts that had been shipped to the United States as Vietnamese goods. The case number is 8282. Interim measures followed while the agency decided, which in practice means the importer’s entries are held and the duty is assessed at the higher rate until someone proves otherwise.

The golf cart is a useful object to start with because it is simple. A cart is a chassis, a motor, a controller and a set of batteries. It can be assembled almost anywhere in a shed with a jig and four workers. If assembly alone conferred origin, every cart in America would be Vietnamese by now. It does not, and the case exists to establish that point at somebody’s expense.

What follows is an attempt to price the three routes a Chinese manufacturer can take in 2026, using published costs rather than intentions. The three are: route the box through a third country, build the plant, or stop selling the finished good and sell the machine that makes it. Each one works. Each one fails at a specific point, and the points are not the ones the trade press usually names.

The word that hides the decision

Chinese policy language uses three different terms that English trade coverage collapses into one. They describe different operations, with different costs, and mixing them is why so much commentary about Chinese factories “leaving” talks past itself.

• Capacity relocation, channeng waiqian, means moving the ability to produce: lines, tooling, process knowledge, suppliers.
• Order relocation, dingdan waiqian, means the buyer places the order with a factory in another country. The Chinese plant keeps its lines and loses the volume.
• Process relocation, gongxu waiqian, means one step in the sequence moves abroad while the rest stays. Final assembly is the usual candidate, because it is the least capital intensive and the most visible.

Only the first of these is a factory leaving. The second is a customer leaving. The third is a step leaving, and a step can leave without any factory moving at all. Most of what has happened to Chinese manufacturing since 2023 is the second and third, which is why the factory count at home has not fallen the way the headlines imply.

The distinction matters for cost, because the three routes below are not alternatives to each other. They are alternatives to a specific step. A company that relocates final assembly has not relocated capacity; it has relocated a process, and it now carries two plants, two management teams and one set of customers who can move again.

Why the three familiar fixes stopped working

Before pricing the routes, it is worth clearing away the three answers that Chinese exporters have used since 2018, because all three have had their prices reset in the past eighteen months.

The first fix was to route the container through a third country and relabel. That route is now the most expensive one available. The United States applies an additional 40% duty under HTS heading 9903.02.01 to goods it determines were transshipped to evade country-specific tariffs, and the measure carries no mitigation or remission provision. It was created alongside the US-Vietnam tariff framework in July 2025 and reissued under Section 301 authority after the Supreme Court ruled on 20 February 2026 that the reciprocal tariffs could not rest on emergency powers.

The second fix was to dual-source: keep the Chinese plant for volume and qualify a second supplier in Vietnam or Mexico for the orders that need non-China origin. This works, at a price. A qualified second source costs money to audit, holds inventory in two countries, and usually produces at a higher unit cost for the first two years. It also does not remove Chinese content from the bill of materials, which is exactly what a substantial transformation test examines.

The third fix was to wait for the tariffs to normalise. The record of 2025 and 2026 argues against waiting. The emergency-powers programme was struck down in February. The Section 122 baseline of 10% expired at its statutory 150-day limit on 24 July 2026. The transshipment penalty survived both events by being reissued under a different statute. What normalisation looks like in this environment is not removal; it is re-labeling under a new legal basis, with the same number attached.

There is a fourth development that changed the arithmetic more quietly. On 13 August 2026 the White House Office of Trade and Manufacturing Policy published a report on transshipment and confirmed that a screening system, referred to in the report as Detective Border, would score shipments before clearance rather than in a post-entry audit a year later. Full rollout is targeted for the end of 2026. The report’s estimate of illegally transshipped value ranges from USD 40 billion to USD 303 billion a year depending on methodology, with a central figure near USD 75 billion and lost tariff revenue of USD 19 billion to USD 26 billion. The range is wide enough that trade lawyers have contested it. The enforcement posture behind it is not in dispute.

Path one: route the box through a third country

This is the cheapest path to start and the most expensive to be caught on. It needs a freight forwarder, a cooperative factory willing to issue a certificate of origin, and about six weeks of calendar time from decision to first sailing.

The cost structure is simple enough to write on one line, which is part of its appeal to companies that do not want to move a factory. A shipment valued at USD 200,000 that is found to be transhipped attracts the 40% penalty, or USD 80,000, on top of whatever duty applies to the true origin. If the product carries a 15% gross margin, the penalty is 2.7 times the entire margin on the order. Add the civil penalty exposure under 19 U.S.C. 1592, which scales with culpability and can reach the value of the goods in a fraud finding, and the expected loss exceeds the order value.

The part that Chinese exporters consistently underestimate is who pays. The importer of record is liable, not the factory. In the golf cart case the importer is an American company that bought in good faith. That company now has a customs file, a hold on entries and a reason to change suppliers. The penalty may be assessed on the goods, but the commercial damage lands on the relationship.

Path one therefore has a strange cost curve. It is profitable for as long as nobody looks, and catastrophic the first time somebody does. With pre-clearance screening arriving at the end of 2026, the expected duration of “nobody looks” is shortening, which raises the effective cost of every shipment in the pipeline before the system is even switched on.

Path two: build the plant

The second path is the one that gets announced. It is also the one whose cost is most often quoted as a headline number rather than a unit rate, which makes it hard to compare with anything.

A workable benchmark comes from the electric two-wheeler sector. Yadea built a plant in Bac Ninh province, Vietnam, with first-phase investment above USD 100 million and annual capacity of about one million units. That works out to roughly USD 100 of capital per unit of annual capacity, before working capital, before the local content programme, and before the two years of management attention that a greenfield plant consumes.

Apply that rate to a mid-sized components exporter moving 200,000 units a year and the capital requirement is about USD 20 million. That is not a large number for a company with a 15% margin on USD 60 million of revenue. It is a large number for the same company when the plant has to be staffed, qualified by the customer, and certified to the same quality standard as the Chinese original, all while the Chinese plant keeps running to pay for it.

The aggregate numbers show how much of this is happening. China’s outward direct investment flow reached USD 213.58 billion in 2025, up 11.1%, with a stock of USD 3.4 trillion, or 7.4% of the global total. Chinese firms now operate 58,000 enterprises in 189 countries, employing 4.761 million people abroad, of whom 2.965 million are local hires, according to the 2025 statistical bulletin published on 23 September 2026. Belt and Road destinations took USD 46.05 billion, or 21.6% of the flow.

Those figures need one clarification that the coverage usually skips. MOFCOM’s January 2026 press conference reported 2025 outward direct investment of USD 174.38 billion, up 7.1%. The September bulletin reports USD 213.58 billion, up 11.1%. The USD 39 billion gap is the financial sector, which the January 2026 figure excludes by design. Two numbers, two definitions, one year. Anyone comparing the 7.1% growth rate with the 11.1% one is comparing non-financial investment with total investment.

Path three: sell the machine instead of the good

The third path is the one that has quietly become the largest, and it is the one that comes closest to solving the origin problem rather than dodging it.

If the customer’s plant is in Vietnam, the machine that fills it can still be Chinese. A Chinese-made industrial robot shipped to a Vietnamese factory is a Vietnamese capital import, not a Chinese consumer good, and no substantial transformation test applies to it in the destination market. The tariff exposure moves to a different line, usually a lower one, and the Chinese seller keeps the technology, the service contract and the spare parts annuity.

The numbers say this is already the dominant response. China’s industrial robot exports rose 48.7% in 2025, and exports exceeded imports for the first time, making the country a net exporter of industrial robots, according to General Administration of Customs data. China installed 295,000 industrial robots in 2024, or 54% of the world total, and domestic suppliers held 57% of the home market, per the International Federation of Robotics’ World Robotics 2025 report.

The same shift is visible downstream. In March 2026 a Chinese automation supplier shipped an entire automobile welding line to a Vietnamese factory rather than the cars it would build. Foxconn’s subsidiary filed an environmental assessment in April 2026 to build humanoid industrial robots in Bac Ninh, with production scheduled from November 2026. STEP Electric is registering a local operating entity in Vietnam and describes the change as moving from following its customers to steering its own expansion. JAKA Robotics has offices in Japan, Germany, the United States and Malaysia, and a plant in Nagoya.

Table 1: Three routes out of the tariff problem, costed on published figures

Route	What it buys	Cash cost	Time to first shipment	What it does not fix
Third-country routing	A different origin label	USD 80,000 penalty on a USD 200,000 shipment, plus duties	Six weeks	Chinese content in the bill of materials
Overseas plant	Real origin, real capacity	About USD 100 per unit of annual capacity	18 to 30 months	Upstream dependence on Chinese components
Equipment exports	A different tariff line	Freight plus commissioning	Three to nine months	The customer’s ability to buy the same machine elsewhere

 

Source: derived from HTS 9903.02.01 penalty terms, Yadea’s disclosed Bac Ninh investment and capacity, and General Administration of Customs export data for 2025. Time to first shipment assumes a leased building rather than a greenfield site.

What the equipment route costs

The equipment route is not free, and its costs are structural rather than financial.

The first is market size. If China installs 295,000 robots a year and that is 54% of the world total, the global market is roughly 546,000 units and the market outside China is roughly 251,000. The export opportunity is smaller than the home market it is meant to replace. A supplier that shifts from domestic volume to export volume is trading a large pool for a small one at a better margin.

The second is customer concentration. A machine tool or robot line sold to a single Vietnamese plant ties the supplier to that plant’s fortunes for a decade, because the service contract and the spare parts run with the installed base. The Chinese supplier has replaced many small domestic buyers with a few large foreign ones.

The third is the one nobody puts in the brochure. The customer is learning to make the product. Installing a welding line in Vietnam transfers process knowledge to the people who will operate it, and the second line they buy may not be Chinese. The equipment route monetises the transition, and it also accelerates it.

Where each route stops working

The useful part of a cost comparison is the boundary, not the ranking. Each of these three routes has a condition under which it stops being rational, and the conditions are specific.

Third-country routing stops working when the product’s Chinese content exceeds what the destination’s origin rules will tolerate. For a product whose value in 2026 is 70% Chinese components and 30% Vietnamese assembly, no amount of documentation will produce a defensible Vietnamese origin. The route is available to assemblers of low-content goods and closed to makers of high-content ones, which is the opposite of how it is usually described.

An overseas plant stops working when the customer will not pay the cost of the second source. A plant in Vietnam, built in 2026 with 60% of the capital productivity of the Chinese original, produces a unit cost that is higher by a margin the buyer has to absorb. If the buyer is a Western retailer operating on 30% gross margins, it will absorb some of it. If the buyer is a distributor reselling at 8%, it will not, and the order goes to whoever is cheapest regardless of origin.

Equipment exports stop working when the destination market is large enough to justify local machine building. That is already happening in Vietnam, where Foxconn’s humanoid robot line and a growing cluster of local automation firms are building the capability that made Chinese equipment cheap in the first place. The window for exporting machines is the window before the customer learns to build them, and in Vietnam that window is closing at a measurable rate: every line installed trains the technicians who will eventually specify the next one.

Choosing between them

For a company facing this decision in the fourth quarter of 2026, the choice is less about strategy than about three measurable facts.

• What share of the bill of materials is Chinese? Above roughly 50%, routing is not available and the choice is between a plant and equipment.
• Who is the importer of record? If it is the customer, that customer is carrying the penalty risk and will demand a plant. If it is your own trading arm, you are carrying it.
• What is the product’s capital intensity per unit of annual capacity? Below about USD 50, a plant is affordable at mid-size volumes. Above USD 500, it is a strategic bet that needs a customer contract behind it.

A practical test that has emerged among suppliers doing this is to ask the buyer for a letter of intent covering at least 60% of the new plant’s output for five years before committing capital. Buyers resist it, because a letter of intent is a liability on their side and a plant is a liability on yours, and the negotiation over which liability comes first is the real content of any relocation decision. There is a fourth fact that does not appear on any balance sheet, which is whether the customer intends to stay. A plant built for a buyer who then dual-sources anyway is a stranded asset with a quality certificate. The Chinese manufacturers that have done this well signed capacity offtake agreements before breaking ground, and the ones that have done it badly built first and negotiated later.

Data basis

Data as of October 2026. Manufacturing output and share of global value added come from World Bank series for 2026. Outward investment figures come from two MOFCOM releases with different scopes: the January 2026 press conference reporting non-financial outward direct investment of USD 174.38 billion, and the 2025 statistical bulletin published on 23 September 2026 reporting total flow of USD 213.58 billion. The difference is treated here as financial-sector investment and is not independently verified. Robot installation and market share figures come from the International Federation of Robotics’ World Robotics 2025 report for calendar 2024, and export growth from General Administration of Customs data for calendar 2025.

Three figures are calculated rather than quoted. Capital per unit divides Yadea’s disclosed first-phase investment by disclosed capacity, ignoring working capital and local incentives. The global robot market divides China’s 295,000 installations by its 54% share, so a one-point error in the share moves the total by about 10,000 units. The penalty multiple divides the 40% duty on a USD 200,000 shipment by an assumed 15% margin.

No exchange-rate adjustment was needed for the 2025 and 2026 figures.

Limitations

• The 15% gross margin used in the penalty arithmetic is an assumption; at a 30% margin the penalty equals 1.3 times the order margin, not 2.7 times.
• Yadea’s capacity figure is a target rather than a verified output, so capital per unit is a floor.
• The export growth rate covers all industrial robots, not the mix between six-axis and SCARA units.

Sources

• Ministry of Commerce , 2025 outward investment bulletin, 23 September 2026
• Ministry of Commerce , press conference transcript, 22 January 2026
• US Customs and Border Protection , EAPA case 8282 and HTS 9903.02.01, 2026 (site home)
• Office of Trade and Manufacturing Policy , transhipment report and Detective Border, 13 August 2026 (site home)
• International Federation of Robotics , World Robotics 2025, industrial robot installations for 2024 (site home)
• Jiemian News via Toutiao , robot makers push overseas, 9 March 2026

 

This article was prepared by the expert team at Global Infinity Labs, drawing on insights and data provided through its MCP data services.

About Global Infinity Labs

Global Infinity  Labs is an AI innovation company specializing in MCP data services.

Email: support@globalinfinitylabs.com

source https://newsroom.seaprwire.com/press-releases/finance/chinas-factories-are-not-leaving-the-last-three-steps-are/

Philippine Tourism Has Three Arrival Counters and a 300,000 Gap

The same eight-month task, run through eTravel, the merged immigration series and a private estimate, yields three totals and two growth rates.

SeaPRwire – On 4 September 2026, the Philippine Department of Tourism gave reporters a number: 4,115,860 international visitors for January to August, up 3.7% on the same months of 2025. On 18 September 2026 the department’s own dashboard displayed 4,415,860 for the identical period, up 0.78%. The distance between the two September 2026 releases is exactly 300,000 people, or 7.3% of the smaller figure. Neither release was withdrawn. Neither was labelled a correction.

For anyone outside the tourism trade this looks like a clerical problem. It is not. It is what happens when an official statistical series changes the instrument that produces it while keeping the name, the headline format and the political purpose. The country still received a quantity of foreign visitors between January and August 2026. Three separate counting systems were pointed at that quantity, and they returned three different answers.

This piece treats the three systems as instruments and runs the same task through each one. The task is simple enough to repeat: count every person admitted to the Philippines between 1 January and 31 August 2026 who is not a resident, break the total down by country of passport, and compare it with the same count for 2025. What follows is the output of each instrument, the arithmetic that separates them, and a short guide to which counter is fit for which job.

One task, stated so that it can be repeated

Precision about the task matters, because most of the disagreement in Philippine arrival statistics comes from people using the same words for different operations.

The unit is a person, not a trip. A visitor who enters twice in eight months counts twice, which is how almost every tourism authority in the world counts. The territory is the customs and immigration boundary, including airports and seaports. The window is 1 January to 31 August 2026, the period DOT reports as the year to date. The classification key is country of passport, which is what produces the familiar table of source markets.

The task has one property that makes it unusually hard to do cleanly. The Philippines has a very large diaspora, and a large share of the people crossing its border on a foreign passport are former citizens coming home to visit family. Under the Balikbayan programme, dating from Republic Act 6768 and amended by Republic Act 9174, a former Filipino who has naturalised abroad can enter visa free for a year. That person is a foreign passport holder on arrival and a returning family member in every other sense. Whether they enter the count decides whether the American market is 818,318 or 926,575.

DOT’s published total for 2025 was 6.44 million visitors, with tourism receipts of PHP 694 billion and tourism direct gross value added of PHP 2.27 trillion, or 8.1% of GDP. In 2025 the sector supported 7.7 million jobs. Those are the numbers the department is accountable for. The arrival counter is the input that feeds them.

Instrument one: a form filled in before landing

The first instrument is eTravel, the government’s online arrival registration. It is mandatory for every inbound traveller, including Filipino passport holders and crew, and it opens 72 hours before departure. A completed form produces a QR code that airlines check at check-in and immigration officers scan on arrival.

Because eTravel is a registration system, its raw output is a count of registrations, not a count of admissions. Almost everyone who registers does land, so the two numbers are close. They are not identical. A registration can be filed and the flight missed; a connecting passenger can be rebooked; a seafarer can register for a leg that is later cancelled.

The 4 September release used eTravel alone. Its market table read: United States 818,318, South Korea 727,379, Japan 350,191, China 310,088, Australia 234,156, Canada 231,972, Taiwan 154,957, United Kingdom 132,370, Singapore 126,812, India 77,883. In the January to August 2026 window, growth was led by China at 69.15% and India at 32.20%, with South Korea down 17.42% and Taiwan up 9.63%.

Instrument two: the form plus the immigration record

The second instrument merges eTravel registrations with records held by the Bureau of Immigration, the agency that actually admits people at the port. DOT launched a public dashboard on 16 September 2026 that runs on this merged basis, and the 18 September coverage described the source plainly: eTravel records from the Department of Information and Communications Technology, plus Bureau of Immigration records.

The merged total for the same eight months is 4,415,860, against 4,381,585 in 2025, a gain of 34,275 people. The market table rearranges itself. In the September 2026 release the United States rises to 926,575, or 20.98% of the total. South Korea sits at 744,732 in the same table, or 16.86%. Japan appears at 321,545, or 7.3%.

Table 1: Three published counts of the same eight-month task, with each publisher’s own stated change

Instrument	Publisher and date	Jan to Aug 2026	Change as published
eTravel registrations only	DOT, 4 September 2026	4,115,860	+3.7%
eTravel plus immigration records	DOT dashboard, 18 September 2026	4,415,860	+0.78%
Private estimate	LPC, reported 7 October 2026	about 4.40 million	roughly flat

 

Source: DOT , Philippine Tourism Data Dashboard, September 2026 (site home); Economic Daily News , 4 September 2026; Migrant Times , 19 September 2026

Instrument three: a number with no published method

The third instrument is a private estimate. On 7 October 2026, Chinese-language Philippine coverage cited LPC, a property research house, putting arrivals for January to August at about 4.40 million, essentially flat on 4.38 million a year earlier. The same report said growth peaked at 12% in February 2026, stalled from April, and turned negative year on year from May. For January to August 2026 it put China at about 303,000, up roughly 93%, and Korea at 744,000, down 22%.

The LPC number is worth including precisely because it is not an official series. It reaches the same order of magnitude as the merged count through a different route, and its market detail agrees with the merged series on Korea while agreeing with the eTravel series on nothing else. But no methodology is published: not the sample, not the treatment of balikbayan, not whether the count is by port of entry or by first destination. A number without a published method is a signal rather than a measurement, and it should be cited as one.

The 300,000 gap does not behave like an addition

The obvious explanation for the gap is that the merged series adds returning overseas Filipinos to the eTravel count. Three hundred thousand extra people over eight months, in a country with ten million citizens abroad, is a plausible size for that flow. The explanation is probably right in aggregate and wrong in detail.

If the merged series were eTravel plus an extra population, no market could shrink when you move from one to the other. Japan shrinks. The eTravel series has Japan at 350,191; the merged series has it at 321,545, a fall of 28,646, or 8.2%. China also falls, from 310,088 to 303,000, down 2.3%. The United States rises by 108,257, or 13.2%. Korea rises by 17,353, or 2.4%.

A series that is a strict superset cannot subtract 28,646 Japanese visitors. Something else is happening. The two instruments are not nested; they are differently shaped. A registration form records the passport and itinerary declared before travel. An immigration record reflects what the officer admitted, under which category, at which port. Reclassify a group, deduplicate a set of records, or change how a transit passenger is booked, and individual markets move in both directions while the total still rises.

That is the useful finding here, and it is more interesting than the headline gap. The 300,000 is not a pile of extra people sitting on top of the first count. It is the net of additions and subtractions across a much larger set of records, which means the market-level numbers from the two series cannot be mixed. A sentence that pairs the merged total with the eTravel country table is not approximately right. It describes a population that no instrument measured.

Four pairings, four growth rates

The growth rate is where the instrument choice becomes politically visible, because 2026 is the year the Philippines is trying to defend a target.

DOT’s stated goal for 2026 is between 6.4 million and 6.8 million international visitors, with 7 million described in September coverage as a stretch figure. Against a 2025 outcome of 6.44 million, the low end of that range is flat and the high end is a 5.6% gain.

Now count the ways to compute a year-on-year change for the first eight months. The eTravel series compares like with like and reports +3.7%. The merged series compares like with like and reports +0.78%. Cross the two and the answer changes sign.

Table 2: The same eight months, four ways of pairing a numerator with a baseline

Numerator	Baseline	Implied change
4,115,860 (eTravel, 2026)	3,969,000 (eTravel, 2025, implied)	+3.7% as published
4,415,860 (merged, 2026)	4,381,585 (merged, 2025)	+0.78% as published
4,115,860 (eTravel, 2026)	4,381,585 (merged, 2025)	-6.1%
4,415,860 (merged, 2026)	3,969,000 (eTravel, 2025, implied)	+11.3%

 

Source: derived from the three series in Table 1; the 2025 eTravel baseline is implied by dividing 4,115,860 by 1.037 and is therefore an estimate, not a published figure

The third row is the one that should worry the department. If a reader takes the newer 2025 baseline and the older 2026 numerator, Philippine arrivals fall 6.1% in the first eight months. If a reader takes the older baseline and the newer numerator, they rise 11.3%. The two published figures are both defensible. The two unpublished combinations are both nonsense, and both are one careless citation away from print.

The 2025 baseline moved by roughly 412,000 people between the two releases, which is larger than the 300,000 gap in 2026. That restatement is the part nobody has explained in public. A statistical office restating a baseline by 9.4% in 2026 is entitled to do so, but it normally publishes a note saying what changed. Without that note, the 2025 comparison set is a moving floor.

What each counter is good for

None of this makes any of the three instruments wrong. It makes them specialised. The mistake is treating them as interchangeable, and the fix is a short routing table.

• For a headline about foreign tourists, use the eTravel series. It excludes most returning citizens by construction, which is what the word tourist is usually meant to capture.
• For airline capacity, airport slots and hotel pipeline planning, use the merged series. It includes every admission, including the balikbayan who occupies a seat and a room.
• For tourism receipts and GDP contribution, use neither. Use the Philippine Statistics Authority’s tourism direct gross value added and DOT’s receipts series, which count spending rather than heads.
• For the 6.4 million target, ask which series the target is measured on. The published target does not say, and the two candidates differ by 7.3% at the eight-month mark.

That last gap matters arithmetically. To reach 6.4 million on the eTravel basis, the Philippines needs about 2.28 million arrivals in September to December, or 571,000 a month. On the merged basis it needs about 1.98 million, or 496,000 a month. The four-month run rate implied by the merged series is already close to the lower threshold; on the eTravel series it is not.

The 6.4 million target is a fourth instrument

A target is not a measurement, but it behaves like one once it enters a press conference. The 2026 target carries a complication: it was set before either the eTravel or the merged series existed in its current form, so the baseline it was built on is neither.

DOT’s own mid-year communications show the strain. On 6 October 2026 the tourism secretary was reported expressing confidence in the 6.4 million goal, with cumulative growth of 3.46% to that date. That 3.46% figure, reported in October 2026, is consistent with the eTravel series and inconsistent with the merged one. By 18 September 2026 the dashboard series showed growth of 0.78% for eight months. Both statements are true of their own instrument. Only one of them can be true of the target, and the target does not specify which.

There is a second complication, which is the China market. The Philippines opened 14-day visa-free entry for Chinese citizens in January 2026 at Manila and Cebu, and DOT has been seeking to extend it to Clark, Caticlan, Bohol and Palawan. Chinese arrivals rose from 1.74 million in 2019 to a peak that has never been recovered, then collapsed, and are now 310,088 on the eTravel basis for eight months of 2026, up 69.15% year on year. The October 2026 LPC estimate puts the gain at 93%.

A market growing at 69% or 93% from a small base is exactly where instrument differences bite hardest, because the base is small and the reclassification is large. If the visa-free expansion proceeds, the 2027 comparison will be built on a 2026 base that two instruments disagree about by 7,000 people in one market and 300,000 in total.

Receipts are the counter nobody argues about

Amid three arrival counters there is one number that has not been restated: money.

Tourism receipts for 2025 were PHP 694 billion. Tourism direct gross value added was PHP 2.27 trillion, or 8.1% of GDP, with shopping, accommodation and food services the largest components. Those figures come from the Philippine Statistics Authority’s tourism satellite account, which is built from expenditure surveys and business records rather than from a form filled in at the airport.

Dividing receipts by arrivals gives a rough yield per visitor. For 2025, on 6.44 million arrivals, PHP 694 billion works out to about PHP 107,800 per visitor, or roughly USD 1,900 at an exchange rate near 57 pesos to the dollar. That is a high figure by regional standards and it is one reason the arrival count matters less than it appears: a 300,000-person disagreement moves receipts by about PHP 32 billion, or 4.6%, if the marginal visitor spends the average. It does not move the GDP contribution, because the satellite account does not divide by an arrival count.

The distinction is worth holding on to. Arrival counters are an input to marketing and aviation planning. Receipts and value added are the economic output. When the two disagree, the output number is the one that has been through an accounting process.

What would settle this

Three things would close the gap, and all three are cheap.

First, publish the bridge. DOT has both series. A one-page table showing, for each source market, the eTravel count, the immigration count and the difference, would explain the 300,000 in an afternoon and end the guessing. Statistical agencies do this routinely when they change a collection method; the silence is a choice, not a constraint.

Second, restate 2025 on both bases and say so. The 2025 figure moved by about 412,000 between releases. Publishing both baselines, with the date each was computed, would let analysts compute a like-for-like change without inferring one.

Third, state which series the 6.4 million target is measured on. This costs nothing and removes an entire genre of misleading coverage, including the genre that quotes a merged total next to an eTravel market table.

Until then, the honest way to cite Philippine arrivals in 2026 is to name the instrument in the same sentence as the number. “4.4 million on the merged series” and “4.1 million on eTravel” are both accurate. “4.4 million visitors, with 818,318 Americans” is not, and it will keep appearing in print because the two numbers are one search apart.

Data basis

Data as of October 2026. The primary metric is international visitor arrivals for 1 January to 31 August 2026, as published by DOT in two releases and by LPC through Philippine media. All three totals are taken as published. The two crossed pairings in Table 2 divide one published numerator by another series’ baseline, which is arithmetic, not measurement.

Three definitions govern the work. The numerator counts persons, not trips. The growth denominator is the same eight calendar months of 2025. Markets are classified by passport, which is why the balikbayan treatment moves the United States line far more than the China line. One adjustment is deliberately not made: no returning-citizen flow is added or subtracted, because no publisher discloses it by market.

Limitations

• The LPC estimate is triangulation only; no sample, method or port coverage is published.
• The 2025 eTravel baseline is derived from a rounded 3.7% growth rate, so it carries about plus or minus 15,000.
• Receipts per visitor divides spending by an arrival count from another collection system, so it is an average, not a marginal yield.

Sources

• Department of Tourism , Philippine Tourism Data Dashboard and 2025 year-end results, September 2026 (site home)
• Economic Daily News , DOT January to August arrivals at 4.11 million, 4 September 2026
• Migrant Times , visitor arrivals edge up to 4.42 million through August, 19 September 2026
• The Nation Week , tourism and statistics, dashboard launch and 2025 value added, 18 September 2026
• China News Service via Toutiao , LPC estimate of 4.40 million arrivals, 7 October 2026
• Philippine Consulate General, Toronto , Balikbayan privilege under Republic Act 6768 and 9174 (site home)

 

This article was prepared by the expert team at Global Infinity Labs, drawing on insights and data provided through its MCP data services.

About Global Infinity Labs

Global Infinity  Labs is an AI innovation company specializing in MCP data services.

Email: support@globalinfinitylabs.com

source https://newsroom.seaprwire.com/press-releases/technologies/philippine-tourism-has-three-arrival-counters-and-a-300000-gap/