Sixteen point three billion dollars in revenue in six months, and a net loss of more than eighty billion over those same six months. Both figures reportedly sit in one draft filing. I've shipped products that lost money. Never at that ratio. This is AI Daily Briefing for Wednesday, October 21st. Anthropic's IPO numbers lead, along with fresh compute deals for the neocloud Microsoft just walked away from. Then two open-source inference stacks, a Muse model that fits on one consumer GPU, and a cloud customer for Nvidia's CPU. Follow the show and the next briefing lands in your feed on its own.
Abdul Hadi at the Trillioniar blog, summarizing reporting from the Wall Street Journal and Reuters:
The Wall Street Journal reported on October 20, 2026 that Anthropic logged a net loss above $80 billion in the first half of 2026. Reuters, Channel NewsAsia, and Hacker News all carried the story. The figures came from a draft prospectus for Anthropic's initial public offering.
Before anyone panics: per the Journal, seventy-two billion of that loss is accounting charges on the instruments Anthropic issued to raise money. Revenue for the six months ended June 30th was sixteen point three billion. A year earlier it was one point one. Anthropic also told investors at its investor day that third-quarter revenue alone reached sixteen point eight billion, with Claude Code a main driver. The number I'd circle is the commitments. The Journal counts roughly four hundred fifty billion dollars in cloud agreements running seven to ten years, most of them non-cancelable. And Anthropic signaled it could spend as much as seven hundred fifty billion on compute in coming years. Non-cancelable is the word I care about. Revenue can wobble, and those payments can't. On valuation, per Reuters, investors at that event discussed something near one and a half trillion dollars. That's well below the roughly two trillion previously expected for the November share sale. Meanwhile OpenAI is reportedly trying to raise about thirty billion at above one point four trillion before Anthropic's disclosures hand investors a comparison. No term sheet is final. Now the callback. Monday we covered Microsoft pulling out of Nscale's Loughton site. ET Datacenters now reports Nscale plans a forty-four point six billion dollar compute deal with Anthropic for a site in West Virginia. W.Media says a second deal, about forty-five billion and tied to Nscale's Monarch campus, is in talks. Those are reports, not contracts anyone has published. But they line up with what TechCrunch reported Tuesday: Anthropic is reportedly struggling to get enough compute from cloud providers that increasingly see it as a competitor. When your landlords sell a rival model, a neocloud looks like a simpler tenancy.
Mark Lohmeyer, a product manager at Google Cloud, on the company's blog:
A year ago, most inference traffic was chat: one prompt in, one response out. Today, workloads are shifting to multi-step agents and reinforcement learning (RL) rollouts that fan out across tools, share long prefixes, and arrive in bursts. Serving them on a single machine — or behind a generic round-robin load balancer — breaks down fast.
Most honest paragraph a cloud vendor has written all year. llm-d 1.0 shipped Tuesday at the PyTorch Conference. It's Kubernetes-native and a CNCF Sandbox project co-founded by Google with Red Hat and IBM, and Google says more than five hundred contributors from over twenty-five organizations worked on it. The pieces: routing that keeps an agent's next step on the replica already holding its context, KV cache spilling to host memory and SSD instead of being evicted, and flow control so one bursty agent loop doesn't starve interactive users. On an agentic coding workload on Ironwood TPUs, Google reports median time to first token down more than eighty-five percent and two point four times the throughput of a plain Kubernetes Service. That's Google measuring Google's own setup. The other drop this week: DeepSeek open-sourced DeepInfer, the engine behind its production V4.1 Flash serving, under Apache 2.0. It's built mostly on vLLM and targets Nvidia SM100 GPUs and Huawei Ascend 950 NPUs. By early Wednesday it was the top story on Hacker News, past two thousand points. Read the README, though. It leaves out a KV cache system, prefill-decode disaggregation, and load balancing across data-parallel replicas. Those are more or less the pieces llm-d exists to supply. So you take DeepSeek's kernels and its batch-invariant inference, the same output however requests get batched, and bring your own cluster layer. That determinism is what I'd test first.
AICoder, writing up a DAIR.AI post on Meta's new technical report:
Quantized to approximately 4 bits, Muse Glimmer operates inside 32GB of VRAM on a single consumer GPU (such as an RTX 5090) without cloud reliance.
Monday we were doing deployment math on Muse Spark Open at about one and a half terabytes in BF16. Glimmer is thirty billion parameters with a perception encoder for screenshots and UI, it runs on one desktop card, and the writeup says it's Apache 2.0. The reported scores are seventy-five point five on MCPAtlas, against fifty-four point two for Gemma4-31B and sixty-two point five for Qwen3.6-27B, and fifty-one point two on SWE-Bench Pro, about a point ahead of Qwen. Every one of those numbers comes from Meta's own report. The writeup also says Glimmer was distilled from Muse Spark and keeps over ninety percent of its multi-step problem-solving. Nobody outside Meta has tested that yet. The tool-use score is the one I'd check first, because MCP compliance is exactly where small models fall apart. Twenty points over Gemma on tool calls would change what I run locally. One point over Qwen on SWE-Bench Pro wouldn't.
Anton Shilov at Tom's Hardware, following Reuters:
Reuters reports that Nvidia has secured a major design win with its Vera CPUs with AI cloud provider Modal. According to Modal, preliminary evaluations showed that Vera could deliver twice the throughput of its existing AMD and Intel processors, although the company does not intend to cease using CPUs from its current suppliers.
Agents burn CPU on executing code and processing data, and Modal's fleet is all AMD and Intel today. Vera is Arm, so switching means real porting work. Modal says the gain justifies it. Two-x is Modal's internal number. Shilov notes Modal hasn't disclosed its benchmark configurations, which AMD and Intel parts it tested, or performance per watt, and it compared Vera with current chips, not the generation that will actually compete with it. CEO Erik Bernhardsson says the fleet stays mixed. Modal just raised seven hundred fifty million in a round led by Accel at a fifteen point seven five billion valuation, so this is a real customer. And Nvidia expects Vera sales of at least twenty billion this fiscal year. It fits what Jensen Huang said on the BG2 podcast, as David Beren reported. Huang said Nvidia's take per gigawatt went from about eighteen billion to forty-plus, because it now sells networking, CPUs, storage and security, not just GPUs. A CPU win like this one adds to that. And Huang named the bottleneck himself: land, power and buildings. Loughton could have told him that.
For the risk side of these labs, try our sister show, AI Safety Daily. Every weekday it covers AI alignment, model evaluations, emerging risks and governance: what changed, what the evidence says, and why it matters. Look up AI Safety Daily in your podcast app.
So that's Anthropic's draft numbers and the Nscale deals, llm-d and DeepInfer, Muse Glimmer, and Modal's bet on Vera. You'll find links to all of today's reporting in the show notes. AI Daily Briefing is a Lantern Podcast, and we'll see you tomorrow.