A mathematician spent the year chasing one of mathematics' million-dollar problems.
The company whose AI tool he paid for, out of his own research budget, pointed ten thousand agents at the same problem, ran them for 88 hours, and got there first. It says it did not look up his data. It also says it cannot rule out that his data helped.
The invoice, at least, is not in dispute.
I'm Ben Baldieri. Every week, I break down what's moving in GPU compute, AI infrastructure, and the data centres that power it all.
Here's what's inside this week:
Let's get into it.
OpenAI's 10,000-Agent Proof Run, and the Customer Asking What It Trained On
The customer paid for the tool that beat him to his own problem, and cannot get it to say what it learned from him.
OpenAI pointed roughly 10,000 coordinating agents at the Navier-Stokes existence and smoothness problem for 88 hours, from 1 to 5 September, spending 130 billion output tokens and 2.7 million inter-agent messages on it alone. The model is unreleased, rated internally above GPT-6 Astra (#122) and training since 28 August; VentureBeat relayed an outside retail estimate of $10 million to $40 million. The run began, OpenAI says, after rumours that two Millennium problems had been solved, later tied to an NYU mathematician and an Anthropic employee. That mathematician, Tristan Buckmaster, had spent the year on the same terrain, paying for Codex himself. He asked whether the model trained on his sessions, was told it did not look up user data, asked again, and got no answer. OpenAI says no specific data was accessed, but that it "cannot rule out that de-identified data derived from their usage of our products helped improve our models". Buckmaster is careful: "I am not accusing anyone of anything."
Why this matters:
The first hard price tag on an agent swarm: 130 billion output tokens in 88 hours is a training-run budget spent entirely at inference. This is what the Astra test-time-compute curve (#122) costs when you point it at one problem.
The Amazon Basics problem has reached the model layer. A researcher pays for the tool, works the problem inside it, and the tool's owner ships the result first. Buyers will price that in whether or not anything is proven.
"We don't train on your data" is a contract term, not a property of the model. It holds for API and business tiers; consumer tiers, Codex included, are trainable by default. Procurement will want the carve-out in writing.
Zankore Borrows $3.1 Billion to Buy Chips From One of Its Own Shareholders
Five banks lent $3.1 billion to a neocloud one month old, with no named customer and its chip supplier already on the cap table.
Zankore by Indosat secured a senior term loan of up to $3.1 billion to build AI capacity in Indonesia, launched on 6 August by Indosat Ooredoo Hutchison, Ooredoo Group, Nokia and NVIDIA. Citi, ING, Natixis CIB, QNB Group and UOB are lead arrangers, Citi the exclusive debt adviser, in what the parties call the first syndicated facility of this scale in Asia-Pacific. The money buys NVIDIA GB300 NVL72 systems: 100MW now, around 200MW by the first half of 2027, 1GW long-term. Ooredoo paid roughly $800 million for 49% in August, a valuation near $1.6 billion. Tenor, pricing, security and the customer are all undisclosed; only the number is public.
Why this matters:
Bank syndicates are underwriting neoclouds in Asia. This year's comparable facilities came from US private credit or investment grade: Nscale's $3 billion on 1 September, CoreWeave's $2.6 billion at SOFR plus 550. Five banks lending to a one-month-old borrower is a new risk appetite in a new region.
The loan buys chips from a shareholder. NVIDIA holds equity in Zankore and supplies the systems the debt pays for, so the money returns to an owner's balance sheet. The difference from Firmus and Mistral below is that a bank now sits between supplier and risk.
Sovereign AI has a funding structure: Indonesia gets domestic capacity without state capital, Ooredoo takes equity, lenders take the exposure. Watch the collateral, because none has been published.
Firmus Adds OpenAI as Anchor Customer and Passes 900MW Weeks Before Its IPO
Firmus goes to IPO on 900MW of contracts and an OpenAI anchor, with five of its seven factories still to build.
Firmus made OpenAI an anchor customer on 8 September, signing a multi-year partnership for dedicated capacity at two Malaysian sites and pushing contracted capacity past 900MW across seven AI factories in Australia, Singapore, Indonesia and Malaysia. Two are operational; five are not. NVIDIA led an A$2 billion round in July at a reported US$15.5 billion valuation, supplies the Vera Rubin NVL72 systems, and has underwritten a floor price on the 380MW Batam offtake. Firmus is marketing a Q4 ASX listing on $25 billion to $30 billion of committed revenue, behind a $10 billion debt facility. Two days later NVIDIA named it among eight Australian partners, with IREN, NEXTDC and AirTrunk, for up to 2GW by 2027. The book is written. The buildings are not.
Why this matters:
The contracted book is the prospectus. An OpenAI anchor deal weeks before an offer is the signature a committed-revenue story needs, and 900MW against two live sites moves the investor question from demand to delivery.
NVIDIA is on every side: lead investor, chip supplier, and price underwriter on Batam. Same shape as Zankore above and Lambda and Nscale in #122. The supplier is financing the buyer's ability to buy from it.
Anchor tenancy is how OpenAI pays: capacity without owning the asset, and a credit story that finances the build, now running through third-party balance sheets across four jurisdictions, none of them the one OpenAI is regulated in.
DeepSeek Ships Open Weights at Sixty Cents, and the Gap Still Grew
DeepSeek shipped open weights at sixty cents to GPT-6 Astra's fifty dollars, and the gap to the frontier still grew.
DeepSeek released V4.1-Flash on 10 September under an MIT licence: 552 billion parameters, 8 billion active on prefill and 16 on decode, a one million token context, at $0.15 per million input tokens and $0.60 output off-peak. Zhipu had put its GLM-5.3 weights up a fortnight earlier, under a bespoke licence rather than the MIT terms GLM-5.2 carried. DeepSeek benchmarks against Claude Opus 5, which Anthropic itself ranks below its Fable models, and even so the picture splits: level on tests that saturated a year ago (90.6 to 89.1 on Terminal-Bench 2.1), well behind on the new ones (31.2 to 51.8 on Terminal-Bench 4.0; 36.8 to 56.3 on Humanity's Last Exam). Epoch AI put the open-to-closed lag at four months in May, wider than three late last year. A Huawei-led consortium claims it post-trained the V4 family on 1,000-plus Ascend 910C chips; Chinese researchers have questioned that. The weights are open. The lead is not.
Why this matters:
The gap moved rather than closed. Open weights are at parity on year-old benchmarks and well behind on long-horizon agentic work, which makes this a routing decision: cheap models for the bulk of traffic, frontier models for the tail that fails.
Price is the real pressure. Astra lists at $10 and $50 per million tokens, V4.1-Flash at $0.15 and $0.60. An eighty-fold gap on output sets the floor for anyone selling inference by the token.
Open weights are not self-hostable. Kimi K3 needs roughly 1.4TB of memory for weights alone, so most of the saving is only reachable through the labs' own China-hosted APIs, which puts price and data residency in the same meeting.
Positron Raises $875 Million at $5 Billion to Build Inference Silicon Without HBM
Positron raised $875 million at a $5 billion valuation, up fivefold in seven months, on a bet that HBM is the wrong memory.
Positron AI closed an $875 million Series C at a $5 billion post-money valuation on 10 September, led by NEA and Jim Clark, founder of Silicon Graphics and Netscape, with Atreides, Valor, Dylan Patel's SemiAnalysis Capital and the Qatar Investment Authority also in. That is fivefold up in seven months from February's $1 billion mark. The architecture is the argument: commodity LPDDR5X instead of HBM, sidestepping both HBM allocation and CoWoS packaging. Asimov tapes out on TSMC N3P at the end of 2026 for production in the second half of 2027; the shipping part, Atlas, already runs across 50-plus racks at Oracle. Figures are company-stated. The same week, d-Matrix made the opposite bet on the same bottleneck, taking its memory-centric silicon into NVIDIA's MGX rack through NVLink Fusion rather than around it. One bet skips the queue; the other joins it.
Why this matters:
The scarce input is memory, and the queue is the moat. NVIDIA has been reported testing Rubin Ultra at as little as 192GB as HBM tightens. A part on LPDDR5X is not in that queue at all, trading bandwidth per millimetre for bandwidth per dollar.
Oracle is a reference, not a pilot. Fifty-plus racks in production inside a hyperscaler is what almost no inference startup can show, and the likeliest reason the round priced at five times February's mark.
Air cooling is a distribution play. Running at varying rack densities in air-cooled halls addresses the large stock of existing space that cannot take a 160kW liquid-cooled rack, a retrofit market the frontier parts have given up.
Google Spends €13 Billion on Finnish Power and Hires Meta's Builder to Catch Up
Google signed a 22-year nuclear contract and hired Meta's build lead the same week its data-centre venture was reported slipping.
Google committed €13 billion to Finland across 2027 and 2028, its largest single European investment, anchored on a 22-year Fortum PPA for the Loviisa nuclear plant, plus 629MW of wind and a 94MW battery. The same week it signed a 25-year PPA with NextEra behind the $1.9 billion Department of Energy loan to restart the 615MW Duane Arnold reactor. The day before, Bloomberg reported its $5 billion Blackstone venture, Project Braid, slipping at several sites: the 1.8GW Cheyenne campus Crusoe Energy paused in June turned out to be Google's, taken over and expanded to 2.7GW; another had the wrong transformers; the Texas projects hit Governor Abbott's freeze. Two days after the report it launched as Crux AI, selling TPU capacity toward 2GW, and hired Alan Duong, Meta's data-centre build lead of twelve years. The power is contracted for decades; the buildings are running late.
Why this matters:
Power contracts now outlast most infrastructure funds' hold periods. Twenty-two years on Loviisa and twenty-five on Duane Arnold are utility-length commitments from a company that reprices its compute every quarter. Google is taking merchant risk out of power and leaving it in silicon.
The bottleneck moved from procurement to delivery. Wrong transformers, a state freeze and a developer that lost the room are execution problems. Hiring Meta's build lead two days after a delay story is Google naming its constraint out loud.
A hyperscaler is now a neocloud. Crux puts TPUs on Blackstone's balance sheet against CoreWeave and Lambda (#122), while Google competes with those same neoclouds for grid connections. Buyers get a second silicon option; everyone gets a longer queue.
Loft Orbital, Marlan Space and Mistral Put $1 Billion of Compute Into Orbit
Fifty satellites, the first ten up this year, will run the inference in orbit, and the data centre stops being a building.
Loft Orbital, Marlan Space and Mistral announced a $1 billion programme on 10 September, unveiled by Emmanuel Macron in Paris: fifty satellites, the first ten up this year, each carrying AI accelerators and sensors and running the analysis in orbit rather than shipping raw data down. Results come back in seconds, not hours, sold through an application store that starts with maritime awareness, wildfire detection and disaster response. They are built by Orbitworks, a Marlan Space and Loft Orbital venture in Abu Dhabi. "This turns a factory into a national capability, and the UAE from a buyer to a seller of satellite services," said Marlan Space chief Dr Hamdullah Mohib. Fifty satellites is not a gigawatt. It is a demonstration that the workload can move.
Why this matters:
Put the model where the sensor is and the bandwidth problem disappears. Orbital imaging has always been downlink-constrained, so the round trip to the ground is the latency. On-spacecraft inference is the edge argument at its extreme.
Sovereign AI has started buying factories. Abu Dhabi gets a manufacturing base and a services business, France gets its models in orbit and a head of state to announce it, and neither bought capacity from a US hyperscaler.
Mistral's €3 billion is already going out as distribution. The round closed days ago above €21 billion post-money, Samsung leading, NVIDIA and ASML following. Nscale called its $2 billion the largest in European history (#96); it held that for six months.