An AI lab took in under five billion dollars last year and lost forty-two billion.
It has signed for five hundred and eighteen billion dollars of computing, about a hundred and thirteen times its sales. Four-fifths of that cannot be cancelled. It has twenty billion in the bank. Its biggest supplier has agreed to lend it money to rent its chips.
None of it adds up today.
All of it is already signed.
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.
Anthropic's Leaked Draft Shows $518bn of Compute, Mostly Locked In
Anthropic has not filed a prospectus, and the draft that leaked already shows what it owes.
Anthropic confidentially submitted a draft S-1 to the SEC in June. Reuters obtained a version and reported it on 28 September. Revenue rose twelvefold to nearly $4.6 billion in 2025. The net loss was $42 billion. About $34 billion of that is an accounting charge tied to future equity. The operating loss was $8.06 billion, up from $2.98 billion in 2024. Cash and short-term investments stood at $20.28 billion. Two unnamed customers supplied nearly a quarter of revenue. In the draft, Anthropic has signed for $518 billion of infrastructure over roughly a decade. The largest are $161.2 billion of Broadcom-related equipment leases, $111.1 billion with Google and $110 billion with Amazon. Microsoft holds $31.4 billion, SpaceXAI (formerly xAI) up to $84.5 billion through 2029 and AMD $20 billion. About 80% cannot be cancelled, Reuters reports. On 25 September Anthropic added $11.6 billion over seven years with Akamai. Then the supplier became a lender. Reuters reported on 1 October that Broadcom agreed to lend Anthropic as much as $42 billion in convertible notes. That could cover about a third of a $125.2 billion, five-year TPU lease. Reuters does not say how that lease sits within the $161.2 billion. Bloomberg reported in August that Broadcom was seeking more than $60 billion of debt for chips for Anthropic and others. On 2 October it reported that banks are about to start selling a $42 billion senior tranche. Anthropic flags the conflict in its own draft. Reuters expects it to become Broadcom's largest compute customer in 2027.
Why this matters:
The suppliers are now the lenders too. Google and Amazon already invest in Anthropic, and Broadcom joins them if its notes convert. Each would sit on both sides of the contract.
Cancellability is the credit term that matters. About 80% of the $518 billion cannot be cancelled. The SpaceXAI capacity is the exception, mostly cancellable on 90 days' notice, so it is the cheapest to walk away from.
Nobody has filed anything. Reuters saw a draft that can change before publication. Several outlets called it a filing this week. It is not one.
AMD Buys World Labs for $8.2bn and Lands Its First Helios Rack Order
AMD spent Monday buying a model company and Wednesday learning who will buy its racks.
AMD agreed on 28 September to acquire World Labs in an all-stock deal valued at $8.2 billion. World Labs, founded by Fei-Fei Li, builds models that generate and reason about 3D environments. Li joins AMD as chief scientist. Two days later HPE announced that Vultr will buy $1.2 billion of AMD Helios racks, its first order for the system. Each rack carries 72 Instinct MI455X GPUs, with HPE Juniper QFX5252 switch trays for scale-up Ethernet. The racks go into Vultr's US data centres. Neither announcement gives a rack count or a delivery date.
Why this matters:
A chipmaker buying a model lab is buying its own demand signal. NVIDIA has spent two years funding customers. AMD is buying one outright, and paying in stock.
Helios needed a first buyer that was not a hyperscaler. Vultr is a neocloud, and its $1.2 billion order shows other neoclouds a second rack-scale option.
AMD sits inside Anthropic's draft as well. Reuters lists a $20 billion computing capacity commitment, more than 16 times the size of Vultr's order.
Firmus Signs Meta Weeks Before Its A$7bn Float
Two frontier labs have signed with Firmus in three weeks, and neither contract has a disclosed value.
Firmus and Meta announced agreements on 29 September for GPU capacity at Firmus AI factories in Southeast Asia. The deal runs on NVIDIA's DSX platform and covers contracted capacity plus options to expand. Reuters reports that the financial terms were not disclosed. Neither company named the country. Firmus operates in Australia and Singapore, with two Malaysian sites in development. Meta already runs GB300 NVL72 systems in Melbourne. Three weeks earlier OpenAI signed for two of those Malaysian facilities. Firmus declined to give that contract value either, or the locations, or the megawatts. It puts total contracted capacity above 900MW, company-stated. The offer opens on 6 October, the prospectus lodges on 8 October and the stock lists on 22 October. Firmus is raising A$7 billion, about $5 billion, and up to $5.5 billion with the greenshoe. That would be Australia's second-largest listing after Telstra in 1997. It also forecasts a $77 million pro forma loss for the first half of FY27. Reuters reported the figure on 24 September, citing two people who had seen the draft prospectus.
Why this matters:
The contract book is the equity story, and it has no numbers in it. Two anchor tenants, two undisclosed values. Buyers are pricing a customer list rather than a revenue line.
$10 billion of Blackstone and Coatue debt sits ahead of the equity. DCD reported the facility in February. Firmus called it one of the largest private debt financings in Australian history. Project Southgate is meant to reach 1.6GW by 2028.
The valuation rests on a forecast. Bankers pitched up to $60 billion, The Nightly reports, which is twelve times the $5 billion of EBIT Firmus targets for FY2028. Most of the factories needed to earn it are not built yet.
OpenAI Prices GPT-6.1 Sol at a Fifth of Its Flagship
OpenAI's new mid-tier model costs a fifth of GPT-6 Astra and, on OpenAI's own numbers, nearly matches it.
OpenAI released GPT-6.1 Sol on 29 September at $2 per million input tokens and $10 per million output tokens. That is a fifth of GPT-6 Astra's standard prices. OpenAI says it nearly matches Astra on agentic coding, computer use and professional work. On OpenAI's figures it scores 6.4 points above GPT-6 Sol on DeepSWE v1.1 and seven points higher on OSWorld 2.0. It also claims a 2.2-point lead over Anthropic's Claude Opus 5.5 on AutomationBench. None of it is independent. The model went live in ChatGPT, Codex and the API the same day. An Ultrafast version promising up to eight times faster generation is still to come.
Why this matters:
Every benchmark here is OpenAI's own, run by OpenAI. Independent results will take a few weeks.
Labs now compete on price per task. Buyers will move work to the model that does most of the job at a fifth of the price.
Cheaper tokens rarely mean fewer GPUs. Anthropic's $518 billion of commitments only work if usage grows faster than prices fall.
CoreWeave Brings Vera Rubin Online, and Cognition Claims 4.8x the Throughput
CoreWeave's first Vera Rubin racks are running production work, weeks after they arrived.
CoreWeave made NVIDIA's Vera Rubin NVL72 systems available on 30 September, NVIDIA said. The first production racks arrived earlier this month. The cluster was running in days. Devin maker Cognition became the first customer to run production work on it. Its early tests show up to 4.8 times the total token throughput of GB200 NVL72 on SWE-2 inference workloads. Each rack also holds 128 Vera CPUs with 11,264 cores for the sandboxes where agents run code. CoreWeave says those sandboxes start more than three times faster. It also launched Forge, which folds Weights & Biases, OpenPipe and marimo into one agent platform.
Why this matters:
The 4.8x figure is Cognition's own test on its own workload. It is also the number NVIDIA chose to lead with, and it will be quoted everywhere.
Agents spend much of their time running code, not writing tokens. NVIDIA now sells the CPUs for that part of the job too.
NVIDIA's Ian Buck said CoreWeave's V100s still run customer work nearly a decade after Volta launched. That is NVIDIA's reply to anyone who thinks GPUs wear out in three years.
AWS Shipped OpenAI, SpaceXAI and Anthropic Models on Bedrock in Two Days
Anthropic's draft lists $110 billion of commitments to Amazon, and Amazon spent the week selling Anthropic's rivals.
AWS made GPT-6.1 Sol generally available on Amazon Bedrock on 29 September, the day OpenAI launched it. The same day it put Bedrock Managed Agents, powered by OpenAI, into preview. It also expanded Claude to India, South Korea and Singapore. A day earlier it added SpaceXAI's Grok 4.7 and Anthropic's Claude Sonnet 5.5, including in GovCloud. That is three rival labs in two days.
Why this matters:
Bedrock is a marketplace, and a marketplace does not pick a winner. Amazon collects $110 billion in compute commitments from Anthropic and revenue from every OpenAI call on Bedrock.
Managed agents are where the lock-in lives. A model can be swapped in one line of code. An agent built on AWS orchestration, identity and data services is much harder to move.
The labs are taking inference to Asia. Claude reached three new Asian markets on Bedrock the same week Meta contracted Southeast Asian capacity from Firmus.
A Solar Developer Pivots to Gas as New Jersey Fines a Site Over 62 Generators
A developer that sold itself as a solar company two years ago is now planning a 450MW gas plant.
Arevia Power is moving into gas and data centres, Heatmap reports. Its subsidiary Rhea Data is developing an Idaho data centre complex on thousands of acres of federal land. The plan includes a 450MW on-site gas plant. Rhea Data is also evaluating a site outside Lincoln, Nebraska. Arevia's chief development officer said solar alone cannot meet the country's growing power needs. New Jersey's environmental regulator fined DataOne $1.07 million for running 62 gas generators without air permits, Marcellus Drilling News reports. Together they total about 123MW. The Vineland site is planned at 300MW and serves Microsoft through Nebius. The generators were temporary backup while Bloom Energy fuel cells went in. In Texas, administrative law judges reviewed a 366MW gas plant for Meta's $10 billion site. They found El Paso Electric failed to consider alternatives. The plant would use 813 modular generators.
Why this matters:
The permits decide the fuel. Heatmap ties the pivot to federal permitting limits on renewables. NextEra and Invenergy have moved the same way.
Temporary backup is where enforcement starts. Vineland's generators were a bridge to fuel cells, and New Jersey fined them anyway. FERC's 50MW netting threshold (#125) and this penalty both catch generation that arrived before the rules.
The Texas judges put the tenant in the room. They said the process was driven by the timeline of Wurldwide, a Meta subsidiary. It is a proposed decision, not a final one.
Everything Else
OpenAI wants at least $30 billion at a $1.4 trillion valuation, Bloomberg reports, a bridge to a 2027 IPO and up from $852 billion last time.
Accelevation's Nasdaq IPO raised $540 million at $18 a share, but 20 million of the 30 million shares came from Olympus Partners affiliates rather than the company.
Samsung committed $1 billion to Helix Digital, the KKR-backed AI infrastructure company whose anchor investors include the Kuwait Investment Authority and NVIDIA.
Lambda closed a $1.008 billion delayed-draw term loan at a 6.78% fixed rate to 2033, rated Baa1 by Moody's and secured on the GPUs it funds and their contracted cash flows.
NVIDIA lifted its buyback by $150 billion on 28 September, leaving $235 billion through fiscal 2028, and Reuters reports it may also anchor Anthropic's listing with as much as $10 billion.
Micron's fiscal fourth-quarter revenue hit $54.2 billion, against $11.3 billion a year earlier, with an 86.8% gross margin and $37.7 billion of net income.
NVIDIA launched an Open Agent Safety Platform on 28 September, pairing its OpenShell agent runtime with a BlueField-4 watchdog, and cited more than 100 partners including Anthropic, Microsoft and CrowdStrike.
Google released Gemini 4 Argon to a small group of cyber partners, claiming benchmark wins over GPT-6 Astra and Anthropic's models.
Crusoe is developing a campus for Google in Armstrong County, Texas, beside the 265.5MW Goodnight 1 wind farm and its sister project under construction.
Lambda will build in Mayes County, Oklahoma, citing $500 million of tax revenue over ten years and paying 100% of its energy costs under the state's ratepayer protection law.
West Java's governor halted BDx Data Centers' 640MW Jatiluhur campus eight days after it broke ground, because its permits have not been issued.
NorthPoint Development offered about 4,500 Hazle Township households $10,000 each, plus $120 million to the township over 15 years, for support of a 1,300-acre data centre project.
A bipartisan Senate permitting deal would make data centres over 20MW pay the full cost of the grid upgrades they need and let states charge them different rates, Heatmap reports.
ElevenLabs ran a $300 million tender at a $22 billion valuation, double February's $11 billion, led by Wellington and T. Rowe Price.