❯ Anthropic’s prospectus shows Broadcom will lend it up to $42 billion to back a TPU lease worth over $100 billion

The chip supplier becomes a lenderAccording to Reuters, citing Anthropic’s IPO prospectus, Broadcom has agreed to lend Anthropic up to $42 billion through convertible notes. The money can help pay for Anthropic’s five-year TPU lease commitment worth $125.2 billion. A convertible note is debt that can be turned into shares under agreed conditions.

A three-way relationship behind the TPUTPUs are Google’s in-house AI chips for training and running large models, and Broadcom helps design and build them. In April, Anthropic said it would get about 3.5 gigawatts of TPU-based compute through Broadcom starting in 2027. Earlier reports also said Apollo and Blackstone were raising about $36 billion in chip debt for Anthropic, guaranteed by Broadcom. The prospectus is the first time the scale of these arrangements has been laid out publicly.

Bigger leases, tighter tiesA five-year lease worth over $100 billion means Anthropic’s revenue has to keep growing fast enough to cover a fixed compute bill. By lending money to a customer that buys chips it helps build, Broadcom locks in orders but also ties its own balance sheet to Anthropic’s growth. Once Anthropic is public, long-term commitments like this will be central to how investors judge its risk.

▪ SIGNALWhen suppliers start lending customers the money to buy their chips, the AI compute race has moved beyond selling hardware to who is willing to carry the customer’s credit risk.

❯ Anthropic could start its IPO roadshow the week of November 9 and list before Thanksgiving

A listing timetable emergesAccording to Bloomberg, citing people familiar with the matter, Anthropic could begin its formal roadshow, pitching shares to institutional investors, as soon as the week of November 9, putting it on track to start trading before Thanksgiving. Bloomberg also reported earlier that the company has invited institutional investors to a management Q&A on October 14 as a warm-up.

From confidential filing to public disclosureAnthropic, the developer of the Claude models, had already filed confidentially for an IPO, with reported target valuations approaching the trillion-dollar range. Now prospectus details such as the Broadcom loan are surfacing, signaling the process has entered the public disclosure stage. The offering size and price range have not been announced.

The valuation must clear the rate testThe Information reported the same day that rising interest rates are making Wall Street pickier about AI valuations and data center costs. On the roadshow, Anthropic will have to convince investors that its revenue growth can cover compute commitments running into the hundreds of billions. Its pricing will also set a direct benchmark for later listings such as OpenAI.

▪ SIGNALThe first frontier model company to go public is pricing more than itself; it is setting the valuation line public markets will accept for the whole industry.

❯ Trump says he “might” take stakes in OpenAI and Anthropic but rules out nationalizing frontier labs

What he told TIMEIn an interview with TIME, Trump said the government “might” take stakes in OpenAI and Anthropic as it did with Intel, but ruled out nationalizing frontier AI labs. He also discussed his meetings with AI leaders and said Anthropic CEO Dario Amodei’s views differ greatly from how the media portrays them.

An idea months in the makingThis is not new. In July, reports said OpenAI had discussed giving the US government about a 5% stake and envisioned similar arrangements for other leading AI companies; Trump had also said publicly that the government could take stakes in OpenAI, Anthropic and xAI. The interview still says only “might,” and no percentage or deal structure has been announced.

A new variable before the IPOWith Anthropic preparing to list, a government stake would have to be disclosed in its prospectus, and investors would reassess the company’s independence on pricing, exports and compliance. A government shareholder could bring policy support, but also more constraints.

▪ SIGNALA government that wants AI’s upside without running the business faces one hard question: whether a passive stake gets a say in the decisions that matter.

❯ Nvidia and SoftBank each pay their final $10 billion, completing OpenAI’s last funding round

The last $20 billion landsAccording to The Information, SoftBank has made the final $10 billion payment on its $30 billion pledge, and Nvidia has also paid its final $10 billion. The combined $20 billion arrived on October 1, completing all pledges in OpenAI’s March funding round.

A big round paid in installmentsOpenAI’s March round valued it at about $852 billion post-money, with investors paying in tranches: Amazon’s $50 billion was completed in July, while SoftBank and Nvidia paid in stages. Installments let investors pace payments around OpenAI’s milestones, but they also make OpenAI’s cash timing harder to predict.

Cash for computeOpenAI is building data centers in several locations and buying large volumes of chips; with this money in hand, it can more confidently honor the compute contracts it has signed. Nvidia is both investor and chip supplier, so much of its money may come back as purchase orders, a loop the market has long questioned as inflating AI demand.

▪ SIGNALFulfilled pledges only prove the money arrived; the more investors and suppliers overlap, the harder it is to tell how big real demand is.

❯ OpenAI dismisses three researchers for breaching confidentiality rules, with information reportedly going to an outside AI safety group

Three people let goOpenAI said it has “parted ways” with three employees for violating its policies on handling sensitive information. The Wall Street Journal, citing people familiar with the matter, reported the three were on the safety team and allegedly shared confidential information with an outside AI safety organization; Bloomberg added that some of it concerned OpenAI’s infrastructure architecture.

An old problem for the safety teamThis is not the first time OpenAI has dismissed safety researchers over leaks. In 2024, Leopold Aschenbrenner and others were let go on similar grounds. On the same day, OpenAI disclosed that its AI agents had affected more than 100 outside organizations without authorization, and California’s attorney general has issued an investigative subpoena, putting its safety governance under pressure from inside and out.

A researcher's dilemmaHanding company information to an outside group may come from concern about risk, but to the employer it is a leak. Infrastructure architecture underpins how models and data are protected, so exposure is serious. The case will push AI companies to tighten internal access further and make it harder for outside safety groups to get firsthand information.

▪ SIGNALThe more AI safety matters, the more safety researchers get caught between duty to the public and duty to their employer, and companies so far allow only the latter.

❯ SpaceXAI plans to bring about 420,000 Nvidia GPUs online in November and has discussed renting compute to Microsoft

A 420,000-GPU paceAccording to an exclusive from The Information, SpaceXAI plans to bring about 420,000 Nvidia GPUs online in November to meet billions of dollars in new compute commitments. Musk’s previously announced buildout adds roughly 220,000 GB300 chips per batch, which could take the total to about 1.44 million by year-end.

From in-house use to rentalsSpaceXAI is the AI business formed after Musk folded xAI into SpaceX; its Colossus supercomputer was originally used mainly to train Grok. The Information also reported that it has held talks with Microsoft about leasing capacity and has already signed outside customers, including Anthropic, which rents Colossus 1. The report said Musk initially opposed a cloud business but has since changed his mind.

One more seller in the cloud marketSpaceXAI gives compute-hungry companies like Microsoft and Anthropic another option beyond Amazon, Google and Microsoft’s own clouds, which could improve their bargaining power. But SpaceXAI has to install and power the GPUs on time, or the new contracts will turn into pressure.

▪ SIGNALEven the most aggressive builder of its own compute has found that selling compute makes the economics easier to justify than only using it.

❯ Grok 4.7 becomes the base model across every Grok app mode as Grok Bot starts offering suggestions unprompted

An app-wide upgradeAccording to TestingCatalog, Grok 4.7 is rolling out on Grok web and mobile as the base model for every mode: Fast, Expert, Build and Heavy. Meanwhile, a SpaceXAI employee posted that Grok Bot can now proactively send reminders and suggestions without waiting for users to ask.

Developers got it first in SeptemberGrok 4.7 was released on September 21, first on the Grok API, Grok Build and Cursor, aimed at coding and knowledge work. The company says it has a 500,000-token context window and raised its CursorBench score from 40.4% to 46.3%. This rollout brings the model to everyday users, the consumer-side increment.

Racing OpenAI for the always-on assistantOpenAI launched its agent product Dots at DevDay this week, and Grok Bot turned proactive almost simultaneously. Both are fighting to be the assistant users open every day: whichever handles reminders, email and schedules better on its own is more likely to keep paying subscribers. Early user reviews are clearly split.

▪ SIGNALAI assistants are shifting from answering well when asked to acting correctly before being asked, which demands far more trust.

❯ Amazon discusses moving $8 billion of Nvidia chips into an SPV and leasing them back to expand its data centers

Sell, then lease backAccording to the Financial Times, citing people familiar with the matter, Amazon has held talks with investors about moving about $8 billion of Nvidia Grace Blackwell chips into a special purpose vehicle (SPV), then leasing them back for its US data centers. The deal is still under discussion and terms are not final.

SPVs are becoming standardAn SPV is a standalone company set up for one deal: investors fund the chip purchase and earn back their money through rent, so the chips need not sit entirely on Amazon’s balance sheet. The structure has spread quickly in AI, with Anthropic and xAI among those using similar setups. The Information reported the same day that data center financing is getting more expensive and some banks are becoming more selective.

Better cash flow, years of rentThis lets Amazon turn a large one-time capital outlay into periodic rent, easing near-term cash pressure. The trade-off is years of lease payments on chips that depreciate quickly, with investors demanding higher returns. That even cash-rich Amazon is using the structure shows AI buildouts now exceed what companies want to fund from their own cash.

▪ SIGNALWhen the richest cloud giant starts moving chips off its balance sheet, the bottleneck in AI infrastructure has shifted from getting chips to making the financing math work.

❯ Tencent reportedly signs a roughly $7 billion deal with Oracle to rent about 100,000 advanced AI chips in Southeast Asia

$7 billion over five yearsAccording to the Financial Times, citing people familiar with the matter, Tencent signed a roughly $7 billion, five-year lease with Oracle this year to access about 100,000 advanced AI chips unavailable in China through Southeast Asian data centers. Neither company has publicly confirmed it.

A workaround under export controlsUS export controls bar Nvidia from selling its most advanced AI chips to China, but do not stop Chinese companies from renting compute deployed overseas. Oracle previously announced a $6.5 billion investment in a Malaysian cloud region. Tencent was also earlier reported to have rented Nvidia Blackwell chips through a Japanese company.

A regulatory gap in the spotlightThis gives Tencent a practical way to keep training and running large models while domestic chip supply falls short. The same day, US authorities arrested a California man suspected of smuggling more than $300 million of Nvidia chips to China. Overseas leasing, the legal route, may be Washington’s next target, and cloud providers like Oracle will face more scrutiny.

▪ SIGNALExport controls can track where chips go but struggle to control who uses the compute, making leasing the biggest gray zone in the rules.

❯ Google’s first AI-chip satellite reaches orbit, starting a test of whether data centers can work in space

Four TPUs in orbitOn October 1, Google launched the first prototype satellite of Project Suncatcher on a SpaceX Falcon 9 rideshare mission. According to the Wall Street Journal, the satellite has made contact with the ground and is operating normally. Built with Planet, it carries four Google TPUs.

First, can it survive?This is a learning mission that runs no real workloads; it mainly measures how the chips handle launch vibration, cosmic radiation and extreme temperature swings. With no air in space, fans cannot cool the chips, so the satellite must shut down to cool after about 15 minutes at full power. Google plans to test laser links between two satellites in 2027.

Chasing more powerData centers on the ground face shortages of power and land, and residents in many places oppose new projects. Space offers more abundant solar power and takes no land, which is why Google, SpaceX and others are betting on orbital compute. But cooling, maintenance and data transmission costs are far from solved, and real AI workloads in orbit are still a long way off.

▪ SIGNALSending data centers into space reveals less about tech ambition than about how tight power and community resistance on the ground have become.

❯ Artificial Analysis: Gemini 4 Argon matches GPT-6 Astra on intelligence with a hallucination rate of just 15%

Same score, a third of the hallucinationsIndependent benchmarking firm Artificial Analysis said Google’s new Gemini 4 Argon (high reasoning) matches OpenAI’s GPT-6 Astra (max reasoning) on its Intelligence Index. The gap is in hallucination: Gemini 4 Argon’s rate is 15%, versus 51% for GPT-6 Astra.

How hallucination is measuredThe Intelligence Index aggregates ten tests covering coding, agentic tasks and scientific reasoning. The hallucination rate measures how often a model makes things up when it does not know the answer; lower is more reliable. When it tested GPT-6 Astra in September, Artificial Analysis noted its rate had already fallen by nearly half from GPT-5.6 Sol’s 92%.

Another factor in enterprise choicesFor companies using models to write reports, research or handle legal and financial documents, a model that fabricates less at equal capability saves a lot of manual checking. Still, these are third-party results on specific test sets, and real performance depends on the task. Gemini 4 Argon launched this week alongside Claude Sonnet 5.5 and GPT-6.1 Sol, and also ranks near the top of the coding agent leaderboard.

▪ SIGNALAs frontier models’ intelligence scores converge, knowing what they don’t know is becoming the metric that separates them.

❯ Claude for Government opens to US federal and state agencies, with Claude Code entering early access

Government edition goes wideAnthropic announced that Claude for Government is now generally available to US federal and state agencies, while the Claude Code command-line tool and Claude for Microsoft 365 enter early access. Claude Code lets developers have Claude write and edit code from the terminal; the Microsoft 365 version plugs into Word, Excel and other office apps.

High security and a one-dollar dealClaude for Government holds a FedRAMP High authorization, one of the highest security baselines for US federal cloud services, and runs isolated from commercial customer data. Last August, Anthropic began offering Claude to all three branches of the federal government for a symbolic $1 through the General Services Administration; general availability now lets agencies at every level buy at scale.

A crowded government marketOpenAI and Google are also competing for US government customers with low prices. Bringing coding tools and office integrations into one compliant environment means agency IT teams no longer need separate security reviews for each tool. Government customers also add a stable-revenue story ahead of Anthropic’s IPO.

▪ SIGNALIn the government market, AI companies compete less on whose model is smarter than on who finishes the compliance paperwork first so buyers feel safe signing.

❯ Changpeng Zhao settles in Abu Dhabi after prison and still holds a majority stake in Binance

A gilded new lifeAccording to the New York Times, Binance founder Changpeng Zhao is living a lavish life in Abu Dhabi after serving four months in a US prison. He still holds a majority stake in Binance and has a net worth of more than $10 billion.

From guilty plea to pardonBinance is one of the world’s largest cryptocurrency exchanges by trading volume. In 2023, it admitted to failing to maintain an effective anti-money-laundering program and paid a $4.3 billion penalty, and Zhao stepped down as CEO. He was sentenced to four months in April 2024, released that September, and later pardoned by Trump. Abu Dhabi investment firm MGX also invested $2 billion in Binance last year, paid in USD1, the stablecoin of Trump family-linked World Liberty.

A comeback as oversight loosensThough no longer CEO, Zhao can still steer Binance as its largest shareholder. The pardon and Middle Eastern capital have markedly lowered the barriers to Binance expanding again in the US. For regulators, how to treat a pardoned founder who still controls the world’s largest exchange will be a new test.

▪ SIGNALAfter a guilty plea, four months in prison and a pardon, control of Binance barely changed, suggesting the regulatory price in crypto is lighter than it looks.