Anthropic's IPO, NVIDIA open-weights, and AI's $36B bet
Claude Fable 5 launched on Tuesday and was gone by Friday. That story ate the news cycle, and for good reason: a government directive pulling a deployed frontier model was a first. But the same seven days had four other structural shifts in the AI industry that are worth your attention, because they will outlast the Fable 5 headlines.
Anthropic filed for a $965 billion IPO
Anthropic submitted a confidential S-1 filing the same week Fable 5 was launching and getting shut down. The valuation target would make it the most valuable AI startup ever to go public. Reported annualized revenue of $47 billion from Claude subscriptions, API usage, and enterprise deals.
The timing says something about the market’s appetite for pure-play AI stocks. Anthropic is betting that the market will price it closer to a platform company than a model provider. The S-1 details are sealed until the SEC clears the draft, but the filing itself signals that Anthropic’s leadership believes the current moment is as good as it gets for locking in a valuation.
Fable 5’s shutdown complicates the narrative. A government can now stop an Anthropic model from serving customers. That risk was theoretical before last week. Now it is a footnote in the IPO prospectus.
NVIDIA released the biggest US open-weights model
Nemotron 3 Ultra landed at Computex on June 1 with 550 billion parameters under a permissive license. It is the most capable US-based open model at that size, and it undercuts the proprietary frontier on price by a wide margin.
NVIDIA has been shipping hardware for AI for years. Releasing a competitive open model at this scale is them sending a signal: the GPU business is not just selling shovels anymore. Nemotron 3 Ultra runs best on NVIDIA hardware, and the inference cost advantage over GPT and Claude makes it attractive for any team that can self-host. The model is shipping now, and the early benchmark numbers put it ahead of GPT-5.5 on several coding and reasoning evals at roughly a third of the output cost.
$35 billion in private debt bought Google TPUs for Anthropic
Apollo and Blackstone structured a $35 billion private credit deal to buy Google TPU chips for Anthropic through a special-purpose vehicle. Broadcom, which co-develops the TPUs with Google, backstopped the senior tranches. Anthropic leases the chips instead of buying them.
This is the largest chip-financing transaction in history, and it matters more than any single model release because it creates the template for how AI infrastructure gets funded going forward. Venture capital rounds cannot cover the compute budgets that frontier training runs now demand. Private credit markets step in, secured by the chips themselves as collateral. Every major lab will look at this structure and replicate it. The financing architecture of AI just permanently changed.
GPT-5.6 was expected any day
Prediction markets gave GPT-5.6 an 80 to 89 percent chance of releasing by the end of June. OpenAI was staying quiet, but the betting was based on the pattern: OpenAI has been iterating on GPT-5 on a roughly 30 to 45 day cycle, and the gap from the last release was approaching the upper end of that window.
The competitive context matters here. Claude Opus 4.8 had just reclaimed the SWE-bench lead a week earlier. NVIDIA released an open model that beats GPT-5.5 on some benchmarks. OpenAI needed to respond, and GPT-5.6 was that response. If it landed in the second half of June, it would be the fourth major model release inside three weeks.
MiniCPM5-1B quietly became the best small model
The story nobody talked about last week: OpenBMB’s MiniCPM5-1B scored 17.9 on the Artificial Analysis Intelligence Index, the highest score ever for a 1-billion-parameter model. It competes with models five to ten times its size.
This matters because small models are where the deployment volume is. Every browser extension, mobile app, and edge device that runs AI on-device needs a model this size. MiniCPM5-1B runs locally, needs no cloud connection, and costs nothing per inference. For anyone building AI features into consumer apps where latency and privacy matter more than absolute capability, this is the model to watch.
FAQ
What is Anthropic’s IPO valuation?
Anthropic filed an S-1 at a reported $965 billion valuation, backed by $47 billion in annualized revenue from Claude subscriptions, API usage, and enterprise deals. The filing is confidential until SEC review.
How does Nemotron 3 Ultra compare to GPT and Claude?
Nemotron 3 Ultra is a 550-billion-parameter open-weights model that beats GPT-5.5 on several coding and reasoning benchmarks at roughly a third of the output cost. It is released under a permissive license and runs best on NVIDIA hardware.
What does the $35 billion chip deal actually fund?
Apollo and Blackstone structured private credit to buy Google TPUs for Anthropic through a special-purpose vehicle. Anthropic leases the chips instead of buying them. Broadcom backstopped the senior tranches. The deal sets a template for how AI infrastructure will be financed.
When was GPT-5.6 expected?
Prediction markets in early June 2026 gave GPT-5.6 an 80 to 89 percent chance of release by June 30. OpenAI did not confirm the timeline, but the release window aligned with its previous iteration pattern.
Why does MiniCPM5-1B matter?
MiniCPM5-1B is the highest-scoring billion-parameter model ever tested, running on-device without a cloud connection. It makes AI features viable in mobile apps, browser extensions, and edge hardware where larger models are too expensive or slow.