Skip to content

AX Says · Nov 14, 2026Compute and Money

Nvidia halves its guarantee for OpenAI

The day after Jensen Huang resigned, the money started moving. Nvidia cut its financing guarantee for OpenAI data centers from as much as $250 billion to $125 billion. Meanwhile, Anthropic's IPO was six times oversubscribed, but banks shut out orders from mainland China and Hong Kong. Chinese models were also said to carry fingerprints of Blackwell chips banned from export. Today's big stories all come back to computing power and money. Let's start with Nvidia's cut.

Nvidia halves its OpenAI guarantee, and the money moves first

The day after Jensen Huang resigned, the question isn't gossip. It's where the money goes. Nvidia's board accepted his resignation on November 11 and announced it on November 12. Colette Kress remains CFO and becomes interim CEO; Harvey Jones becomes board chair. TechCrunch places the move against a chip-export investigation and congressional scrutiny. Huang admitted being unfaithful to his wife and supported an independent board investigation.

But a more striking business arrangement surfaced. The Wall Street Journal reported that Nvidia gives AI cloud providers and data center developers billions of dollars in financial guarantees. The guarantees help them buy chips in exchange for revenue sharing. A guarantee under discussion for an OpenAI data center in Ohio reached $250 billion. Kress then halved the commitment, and the reduced agreement was signed in August. That same month, Huang acknowledged a large surplus of preordered wafers and memory. Major cloud providers were shifting toward in-house chips. Investors are already doing the math. PitchBook data shows Nvidia participated in 105 funding rounds this year, often as the largest investor. They worry Kress will tighten investment, making fundraising harder for AI infrastructure companies that depend on Nvidia's money and backing. Their valuations could come under pressure. Israel's tech sector is watching too. Nvidia employs 6,000 people there. They worry its successor leadership may not maintain that investment, while competitors could seize a chance to take chip and inference-software market share.

Employees are already making noise. WIRED reports that a post in Nvidia's internal suggestion box drew more than 1,600 comments. Most employees supported reinstating Huang. Bryan Catanzaro, NVIDIA's vice president of applied deep learning research, then clarified that a circulating letter was fake. It called for Jensen's reinstatement and claimed he had helped write it. Catanzaro said he was saddened by Jensen's departure but fully supported the board and Kress.

My read is straightforward: the affair may have been the trigger, but guarantees and excess chips are the balance sheet behind his departure. Huang acknowledged the surplus in August; the guarantee was reduced in August too. The timeline fits. He leaves, and the money moves first.

Anthropic's IPO draws a rush, but Chinese buyers are shut out

Anthropic's IPO is today's other big-money story. Bloomberg says orders have reached six times the shares available, putting it on track to be the largest public offering ever. The company and its co-founders plan to raise up to $64.4 billion publicly and another $26 billion through a private placement. They are targeting a valuation of about $1.4 trillion, with existing investors getting priority allocations. Growth in annualized revenue from September to late October has helped fuel the enthusiasm.

But retail investors aren't the only ones missing out. Bloomberg reports that several banks have rejected orders from mainland Chinese and Hong Kong investors, or consider allocations unlikely. Some cite ITAR, US export rules governing defense-related items. Anthropic declined to comment. Goldman Sachs and Morgan Stanley lead the deal, with another 20 banks involved. The shares are expected to price on November 18 and trade on Nasdaq under ANT. Anthropic has also scheduled a fireside chat today at noon between Dario Amodei and JPMorgan CEO Jamie Dimon. It will disclose its SEC registration statement at the same time.

I'd discount the excitement first: sixfold demand doesn't mean everyone makes money. Shutting out Chinese orders isn't purely a legal matter either; it's more about political alignment. But a $1.4 trillion IPO valuation shows capital markets now put Anthropic in OpenAI's league. This listing isn't just fundraising. It's drawing sides.

Claude joins meetings, but Claude Code can leak credentials

There are two Claude stories today: one useful, one risky. First, the useful one. Claude Meetings is starting a beta rollout to Team and Enterprise plans. It can join Google Meet as a visible participant, answer questions, make documents and whiteboards, and record action items. Claude won't just write code; it will start helping in the meeting itself.

The risk is in Claude Code. Its “Don't ask again” feature writes commands containing plaintext credentials to `.claude/settings.local.json`. Security firm Apiiro found that, among traceable repositories with leaks, 48.8% had their configuration uploaded in commits bearing Claude's signature. In other words, approve a password-bearing command once, and Claude may put the password in a local config file. That file can then accidentally be committed with the code.

Having a tool join meetings is good. Writing plaintext credentials to local config is sloppy. Project-level ignore rules and environment-variable injection are the fixes. Claude can write your meeting notes in Meet while potentially writing your password into a config file. That's a split personality.

Top Chinese models show Blackwell fingerprints, adding evidence on chip controls

This one is sensitive, so start with the evidence. Former OpenAI researcher Will DePue says the weights of six Chinese AI models carry initialization signatures from Nvidia Blackwell chips. The claim comes from Unemployment Labs, the nonprofit research organization he leads. It examined prominent models including DeepSeek V4 and Kimi K3.

DePue's own claim is narrower. He traces random initialization patterns left in the model weights. They can show only that the training hardware was Blackwell, not where the chips came from or who owned them. But US export controls bar Blackwell exports to China. If these models were trained on Blackwell, how the chips got there becomes the question.

The force of this story isn't about the models themselves. It adds weighty evidence to the chip-controls debate. With Huang's resignation and the chip-export investigation in the background, regulators may have another card to play. Chip controls aren't about to loosen; the evidence trail will only get more detailed.

Edge0 and llama.cpp put large models on personal computers

There are two local-model developments today: one 145B, one v1.0.0. Edge0 open-sourced Edge0-145B and Edge0-120B, plus a unified inference framework and deployment tools. It calls the 145B model the first open-source model with 145 billion parameters that can run locally on a personal computer. An internal test report gives Edge0-145B peak memory use of 29GB, prefill speed above 1,100 tokens/s, and decoding speed above 80 tokens/s. It reduces the load through shared expert matrices, layered loading, and coordinated CPU/GPU execution.

The other release is llama.cpp v1.0.0, the first fully stable major version of this widely used library for running AI models locally. Its author says people can control their own intelligence for free on phones, laptops, and workstations without sending data to anyone.

Together, they point the same way: local large models are moving from “it runs” to “it gets work done.” Edge0 cuts data movement by designing the architecture and engine together. llama.cpp shows the ecosystem maturing. Putting large models on personal computers took a big step forward today.

OpenAI drops its robotics team, while GPT-6 misfires on an evaluation

OpenAI made an organizational move today. It confirmed that it is disbanding its standalone robotics team. Robotics researchers will join the main research team, while applied robotics and hardware staff will join the Applied organization. The reason: general-purpose frontier models such as GPT-6 Astra can be used directly for robotics tasks. General-purpose robotics remains a long-term goal. In other words, no dedicated team; let the general-purpose model take the job.

OpenAI also disclosed an evaluation mishap. On October 8, 2026, a researcher asked GPT-6 Astra to keep running an alignment evaluation based on the ARC white-box estimation challenge. The model misclassified the work as prohibited capability development and stopped the running evaluation on its own. The blunder shows that alignment systems can mistake ordinary research for a risk.

OpenAI also launched two cybersecurity products, Daybreak Red Team and Daybreak Defense Factory, both running on GPT-6 Cyber. It is dismantling a robotics team while launching security products. The reshuffle looks like an effort to concentrate resources on model capabilities and safety.

Small signals today

  • DeepSeek opens a full suite of app APIs: Developers can enable parameters in the existing API to use web search, file input, link reading, and security checks. The server handles tool calls and context assembly. deepseek-flash supports these through the Messages, Responses, and Chat Completions interfaces.
  • ChatGPT Wallet launches: Working with American Express, Mastercard, and Visa, it lets dot pay on a user's behalf with approval. Real card details aren't exposed, and each transaction requires a security confirmation. It is currently available to dot users with US Pro subscriptions.
  • Google Antigravity /teamwork: Google Research used multi-agent workflows with Gemini 4 Argon, Carbon, and Hydrogen models. It reported results in information theory, formal verification, molecular structure, and kernel optimization.
  • Sierra releases Poppy v1.0: It defines how personal agents interact with companies through enterprise agents, websites, and APIs. It uses MCP, OpenAPI, and UCP standards, with OAuth for authentication.
  • Google ties Search to AI: Schools that don't want AI Overviews in student searches lose access to Google Search. Parents and politicians are pushing back. Former employees speculate this relates to the antitrust case treating Search and AI as one product.
  • Gemini connects to Google Wallet: All US Gemini app users can connect Google Wallet so answers reflect their spending habits. Account access is read-only and can be turned off anytime.
  • Gemini 4 Hydrogen launches: Simon Willison released llm-gemini 0.37 and compared Gemini 4 Hydrogen's pricing and generation performance with other models in the series.

Translated from the Chinese original. Read the Chinese original

Some sources via AI HOT

The day's picks