The biggest tenant on America's biggest AI campus has written to its landlord about the rent.
The state has twice refused to let a pipeline cross five thousand two hundred feet of its land. About twenty banks lent eighteen billion dollars against a lease that assumes the power arrives. The tenant is not leaving; it is reserving the right to stop paying.
The campus was financed on a signature.
Now the signature depends on a pipeline.
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.
Oracle Cites Force Majeure on Stargate's Biggest Site
Five thousand two hundred feet of state land is holding up 2.45 gigawatts, and Oracle has written to its landlord about the rent.
Oracle sent a force majeure notice to Blue Owl Capital, whose Stack Infrastructure unit is building Project Jupiter. The 2.45GW campus in southern New Mexico is one of five US sites under the $400 billion deal OpenAI signed with Oracle and SoftBank. Oracle is not giving up the tenancy. It is positioning to defer rent if the site misses its 2028 date. Bloomberg reports that an agreed force majeure event tied to power could push the start of rent payments back three years, that about 20 banks lent $18 billion against the build, and that Oracle would still pay other costs and still owe the full term. The trigger is gas. An Energy Transfer pipeline due in service in September slipped to 1 February 2027. The Bureau of Land Management cleared the federal 16 miles in May. New Mexico's State Land Office refused rights of way twice, in March and again on 14 July. The two blocked stretches run 3,200 feet in the north and 2,000 feet in the south, on a 17 mile line. On 25 September Bloom Energy said Oracle had reaffirmed its contract for 2.4GW of gas fuel cells at the site, company-stated.
Why this matters:
The lease is the collateral. About 20 banks lent $18 billion against an Oracle covenant, not against New Mexico dirt. The FT reported the loans quoted at 89 to 91 cents six days before the notice landed, so the market had started pricing what underwriting did not.
Watch whether the other anchor tenants follow. Google's termination right on its SpaceX compute deal bites on 30 September if committed GPU access has not arrived, after a month's grace. Oracle has shown the softer version, which is to stay in the lease and move the rent.
Snorkel Triples to $3.5bn on an Eighteenfold Revenue Year
The labs signing those compute contracts are paying a second invoice, and this is the company writing it.
Snorkel AI sells the labs their training and evaluation data, the examples a model learns from and is then graded against. It raised a $350 million Series E at $3.5 billion on 22 September, co-led by Insight Partners and S32, nearly triple the $1.3 billion it carried 16 months earlier. Snorkel says its annualised revenue run rate is $375 million, up eighteenfold in twelve months, company-stated. It moved from selling data-labelling software to producing the data itself, generating synthetic examples alongside human experts rather than running a labour marketplace. Mercor is at $2 billion gross annualised, Handshake at $1 billion and Micro1 at $500 million. All three pay 60 to 70% of revenue to the specialists doing the work. Snorkel books those payments in cost of goods sold, so its figure is not gross-comparable to theirs.
Why this matters:
Training data is repricing as a margin business rather than a headcount business. Mercor, Handshake and Micro1 hand 60 to 70% of revenue to specialists. Synthetic generation is the only way out of that.
Investors marked Snorkel at roughly nine times forward revenue. That is a compute multiple on a company that owns no compute.
Crusoe Signs Thinking Machines at $65m a Year for Inference
Thinking Machines raised money to build frontier models and is paying somebody else to serve them.
Crusoe and Thinking Machines Lab, the frontier lab behind the Inkling models, signed a $65 million annual agreement on 23 September for production inference on NVIDIA HGX B200 systems over Quantum-2 InfiniBand, delivered as a dedicated Tailored Deployment through Crusoe Managed Inference. Thinking Machines will serve its Inkling models, GLM 5.2 and 5.3 and fine-tuned variants on it. Crusoe says Managed Inference has passed $100 million in contracted annual recurring revenue less than a year after launch. It follows the Perplexity deal and Crusoe's own $3.9 billion raise at $30.9 billion (#124). Both companies supplied every figure here.
Why this matters:
Two labs in nine days, both for inference rather than training, both on dedicated rather than shared capacity. Inference contracts are becoming the recurring revenue neoclouds have never had.
An American neocloud serving GLM weights for an American lab is the quiet consequence of China's open-weight strategy. Z.ai funds the models (#124), Crusoe rents the racks.
Basecamp Raises $140m to Point a Foundation Model at DNA
Roche's vice-chairman, NATO's venture fund and NVIDIA all wrote cheques into the same genomics company this week.
Basecamp Research raised a $140 million Series C on 23 September led by S32, with NVIDIA, Menlo Ventures' Anthology Fund, the NATO Innovation Fund, the Rockefeller Foundation and a fund called Sovereign AI among the backers, alongside Roche vice-chairman André Hoffmann. The London and Boston company trains EDEN, a biological foundation model, on its Trillion Gene Atlas, a proprietary dataset built with NVIDIA, Anthropic, PacBio and Ultima Genomics from sampling partnerships across more than 30 countries. The money funds EDEN-designed therapeutics, starting with in vivo cell therapy. Existing cell therapies cost hundreds of thousands per patient. Preclinical results are described as strong across multiple modalities, with no data disclosed.
Why this matters:
The moat is the dataset, not the model. EDEN's edge is sampling rights across 30-plus countries, which no amount of compute reproduces, and it is the same argument Snorkel is making one layer up.
S32 led this and co-led Snorkel the day before. One fund is buying both ends of proprietary training data in the same week, human expertise and biological sequence, while NVIDIA again invests in its own customer.
Alibaba's T-Head Announces Zhenwu V900 and a 20GW Target
Alibaba put a chip, an interconnect and a 20GW build target on one stage, which is the NVIDIA stack described without using the word NVIDIA.
At the Apsara Conference on 22 September, Alibaba's T-Head unit introduced the Zhenwu V900, claiming three times the performance of the M890 with 216GB of memory and 1,200GB/s of inter-chip bandwidth. T-Head says more than 1,000 chips run as a single system over its own ICN Switch, and that one cluster scales to 500,000 accelerators. The Zhenwu series has served more than 650 enterprise customers, and M890 supernodes already run models above two trillion parameters including Qwen3.8 and Kimi K3. Mass production is set for Q1 2027, with Yitian CPUs following in Q3 2027. CEO Wu Yongming put the strategy as models, chips and cloud, and set Alibaba Cloud a target of more than 20GW of data centre capacity worldwide by 2032. Every figure above is T-Head's, and TechNode is the only outlet carrying most of them.
Why this matters:
20GW by 2032 is hyperscaler capex stated as a power number, and it is the figure that makes the silicon credible. A chip roadmap without the megawatts to run it is a press release.
The interconnect is the tell. ICN Switch at 1,000 chips is an NVLink answer, and interconnect is exactly where Huawei's pod came in smaller than promised last week (#124).
Q1 2027 silicon against Vera Rubin shipping now is a real gap, but the customers are captive. 650 enterprises on Zhenwu do not benchmark against Blackwell, they buy what Alibaba Cloud sells them.
Meta's Petal Doubles Transoceanic Capacity to a Petabit
Meta is laying its own transatlantic cable rather than buying space on somebody else's, and doubling the ceiling on the way.
Meta announced Petal on 21 September, a 7,000km cable from the US to France carrying 1 petabit per second, double its Anjana system on Meta's own comparison. Meta says it is the first cable system to deploy multi-core fibre at transoceanic distances. It runs 2-core fibre in a 24 fibre-pair system, equivalent to 48 single-core pairs, and fits inside existing 18kV power feeding equipment rather than demanding more. NEC is turnkey supplier, Sumitomo Electric developed the 2-core fibre and Orange handles the landing on the French Atlantic coast. Service is expected in 2029.
Why this matters:
Subsea capacity has always scaled by adding fibre pairs, and therefore power, and a cable's power budget is fixed at the shore. Multi-core fibre breaks it, the first real headroom in transoceanic bandwidth in years.
Cost per bit on the route roughly halves, and Meta sets that price for itself rather than negotiating it from a consortium. Owning the glass means scheduling your own traffic instead of queueing for someone else's.
Service in 2029 means it is sized for training runs nobody has specified yet. Cables are among the longest-lead items in the stack and the least discussed constraint on spreading compute across continents.
Amazon Commits Up to $8bn to Generac and Takes a Warrant
Amazon is buying its way off the grid, and paying part of the bill in Generac equity.
Generac disclosed a long-term supply agreement with Amazon on 16 September worth up to $8 billion. Initial deliveries are scheduled to total $2.4 billion across 2027 and 2028. Amazon received warrants over 1,693,745 shares at $200.93. Only 307,954 vested on signing, and the rest vest against payments Generac actually receives. Generac stock rose more than 40% in extended trading. Cummins and Caterpillar have seen data-centre backlogs roughly double over the past year. Power-segment revenue grew in the high teens. The demand is a queue problem rather than a resilience one: operators who cannot get grid capacity on a timeline they can finance are generating their own. FERC opened one route in June and closed another. It put show-cause dockets on every major grid operator over co-located load, and set a 50MW netting threshold on PJM's tariff.
Why this matters:
"Up to $8 billion" is a ceiling. The scheduled part is $2.4 billion across 2027 and 2028. Amazon holds the rest as an option, and Generac has not booked it.
The warrant is the interesting half. Amazon is paying for supply security partly in equity upside, and only 307,954 of the 1,693,745 shares vested on signing, so that half is a ceiling too. It is the structure NVIDIA runs with neoclouds, now applied to engines.
The phrase to watch is prime power. Emergency plant that runs continuously needs its own permit, its own fuel contract and its own emissions profile. FERC's 50MW threshold is the first of those rules on paper. The rest are being written now.
Everything Else
Anthropic released Opus 5.5 at $20 per million output tokens, down from $25, and Artificial Analysis scores it 58, top of its Intelligence Index at $5.98 a task.
OpenAI launched GPT-6 Sol and Luna ninety minutes later at half the API price of the GPT-5.6 versions, crediting caching and inference gains.
Xiaomi open-sourced the MiMo-V2.6 series the same day, and MiMo-V2.6-Pro scores 46, top of the open-weight models at $0.13 a task, on a published $2.62 million for the six-day reinforcement-learning run.
Anthropic's wet lab said Claude found a CRISPR-like enzyme system in 21 hours on 950 agents and 210 million tokens, a comparison independent researchers dispute.
Nscale filed a Form S-1 for a New York listing under NSCL, showing $140.6 million of first-half revenue, a $1.02 billion net loss and $103.4 billion of active and contracted total contract value as of 31 August.
CXMT started mass production of its fifth-generation DRAM platform, company-stated at 50% more dies per wafer than the node before it, LPDDR5X parts first.
Lumentum, Qualcomm and Corning demonstrated optical die-to-die interconnect at ECOC, roughly 1 Tb/s per millimetre of shoreline with a stated path to four.
Huawei Cloud set commercial launch dates for its Ascend 950 cluster service, 30 September in China and 30 November globally, at a company-stated 1 EFLOPS FP8 across 1,024 cards.
Texas PUC softened its data-centre interconnection rules, dropping a $50,000 per MW non-refundable fee for a flat $100,000 study fee, stretching the energisation notification window from 6 to 24 months and raising the security refund from 20% to 100%, while the 474GW ERCOT queue stays paused pending audit.
Victoria's new data centre plan requires new sites to bring their own renewable generation and storage, match demand with new supply, use non-drinking water for cooling and sit 150 metres from homes.
Google and Georgia Power agreed nuclear uprates at Plant Vogtle and Plant Hatch adding 96MW, with roughly $900 million of projected customer benefits over the life of the units, pending PSC approval.
Verda, the Helsinki GPU cloud formerly called DataCrunch, raised $189 million led by Emergence Capital, taking company-stated total funding past $450 million.
X-energy advanced its Xe-100 small reactor on federal funding. Centrica has picked the same design for up to 6GW of new British nuclear.