Issue #130: Anthropic Lost $80.7bn in Six Months. Most of It Was on Paper.

The company that buys more compute than almost anyone has finally opened its books.

Revenue is up nearly fifteen times in a year. It has signed four hundred and fifty billion dollars of cloud contracts it cannot cancel. One lender is being offered more than ten per cent to fund its chips. A neocloud that just lost an anchor tenant wants forty-five billion more from it.

Everyone is lending against the same customer.

That customer is still borrowing.

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 draft prospectus shows revenue up nearly fifteenfold and $450 billion of compute it cannot walk away from.

Anthropic's draft prospectus shows revenue of $16.3 billion in the first half of 2026, up from $1.1 billion a year earlier, the Wall Street Journal reported on 20 October. It posted a net loss of $80.7 billion, of which about $72 billion came from non-cash accounting charges on convertible notes and other financing instruments. It spent nearly $9 billion on computing capacity and infrastructure in the half, against $7.3 billion in all of 2025. The filing lists about $450 billion of noncancellable cloud commitments over the next decade, including $111 billion with Google Cloud, $110 billion with AWS and $31 billion with Microsoft, and says spending could reach $750 billion including deals still being negotiated. Two customers made up about 20% of revenue. The listing could value Anthropic at about $2 trillion. Separately, the FT reported on 16 October that Anthropic offered Blackstone a yield of more than 10% to put $9 billion into debt backed by five years of its chip lease payments.

Why this matters:

China's best-known open-weight lab now has serious money, and it is still a fraction of what its US rivals have raised.

DeepSeek raised about 100 billion yuan, roughly $15 billion, Bloomberg reported this week, after lifting its target from 50 billion yuan to 80 billion and then again. Battery maker CATL and Tencent were among the largest investors. No valuation was disclosed. Bloomberg says DeepSeek has increased spending on data centres and open-source models and is preparing for a possible listing, and notes that OpenAI and Anthropic have each raised more than $100 billion. The same week, DeepSeek open-sourced an inference framework that runs on both NVIDIA and Huawei chips, TechNode reports.

Why this matters:

In one week Nscale lost a Microsoft site and heard investors price its IPO at little more than half its ask.

Microsoft has terminated its contract for Nscale's planned Loughton campus in Essex, The Telegraph reported on 18 October, and Microsoft confirmed the termination. The site was due to host 23,040 NVIDIA GB300s for Azure across 50MW, expandable to 90MW, and UK Power Networks reportedly told Nscale it could not supply enough power until the early-to-mid 2030s, Data Center Knowledge reports. Nscale says its other Microsoft contracts are unaffected. Bloomberg has also reported its flagship Glomfjord site in Norway running six months late, with 6.4MW of a planned 30MW live. On 18 October, Bloomberg reported that Nscale wants a valuation of about $30 billion for a New York listing raising up to $3 billion, and that a group of investors put $16 billion to $19 billion as more likely to sell. Its last private valuation was about $15 billion. The prospectus, filed on 18 September, shows $103.4 billion of contract value at the end of August, of which $2.6 billion was active. On 19 October, Bloomberg reported Nscale is close to a roughly $45 billion deal with Anthropic for 460MW of Vera Rubin capacity in a second building at its Monarch campus in West Virginia. Neither company commented.

Why this matters:

Recursive's first model ships with an inference engine written for exactly three Apple chips.

Recursive released Flint on 21 October, a mixture-of-experts model with 40 billion total parameters that activates 2.5 billion per token, with weights on Hugging Face under an Apache 2.0 licence. The company says a small team built it from scratch in under four months, with its Eureka Machine, an automated research system, contributing to data, training recipes, quantisation and kernels. Alongside it, Recursive open-sourced rec, its own inference engine, built on Apple's MLX framework and written for three chips only: the M4 Pro, M5 Pro and M5 Max. The code tunes its Metal kernels separately for each chip and needs at least 48GB of unified memory for the roughly 23GB of 4-bit weights and a 256,000-token context. Recursive says rec runs Flint at 1.4 to 2.6 times the throughput of llama.cpp on an M4 Pro, and at more than 125 tokens per second at short contexts. Across Recursive's own six-test coding average, Alibaba's Qwen3.8 27B still scored higher, TechCrunch reports. Recursive emerged in May with $650 million at a $4.65 billion valuation, backed by GV, NVIDIA and AMD Ventures, and signed a $410 million compute deal with AWS in June.

Why this matters:

China's biggest tech companies ordered more than 3 million NVIDIA cards that still clear US export rules. They may get as little as a tenth of them.

ByteDance, Alibaba and Tencent have ordered more than 3 million NVIDIA RTX PRO 5500 cards at about $11,400 each, or more than $35 billion combined, Reuters reported on 20 October, citing two sources. Total orders from Chinese customers have passed 6 million cards. The RTX PRO 5500 is a workstation card and among the most powerful NVIDIA parts Chinese buyers can still get. The B300 cannot be sold to China, and the H200 needs a licence reviewed case by case. NVIDIA has told customers it cannot guarantee supply and can make no more than 3 million of the cards worldwide next year, as it shifts TSMC capacity to Rubin. One source said partners may receive only 5% to 10% of what they ordered. Samples are due in December, with deliveries from the first quarter.

Why this matters:

Two hyperscalers went after the software habit that keeps AI on NVIDIA, a day apart.

AWS released TorchNeuron on 19 October, a version of PyTorch that Amazon says moves an app onto Trainium with a single configuration change and no code rewrite, TechCrunch reports. It will also contribute code upstream to strip chip-specific dependencies from PyTorch, and claims Deepgram processes 70% more audio per dollar on Trainium. The next day, Google open-sourced TorchTPU, its native PyTorch stack for TPUs, built with Meta's PyTorch team. Google says training on TorchTPU reached up to 52% model FLOPs utilisation across as many as 8,192 Ironwood TPUs, and that Hugging Face models can now target TPUs with a one-line setting. Amazon says about 85% of AI apps are written in PyTorch.

Why this matters:

Nebius is set to anchor a Saudi-backed AI campus on the Red Sea, six weeks after construction began.

HUMAIN, the Public Investment Fund's AI company, and DataVolt announced a preliminary framework agreement with Nebius on 21 October for capacity at a planned AI campus in Oxagon, NEOM's industrial city on Saudi Arabia's Red Sea coast, W.Media reports. Nebius would take up to 50MW, half of the first 100MW now under construction, out of a planned 360MW first phase. The capacity would come online in phases from the fourth quarter of 2027, and Nebius is expected to deploy NVIDIA-based infrastructure there. The announcement came six weeks after the partners said construction had started. No financial terms were disclosed.

Why this matters: