❯ Broadcom’s bank group starts placing $60 billion of AI chip financing, with Anthropic the main beneficiary

$60 billion goes to marketAccording to Bloomberg, citing people familiar with the matter, Broadcom’s Wall Street bank group has begun gathering about $60 billion of AI chip financing that will benefit Anthropic and other companies. The banks are preparing to send investors syndication letters for a $42 billion Class A senior secured portion. The deal has not been formally announced.

Debt in two layersThe financing has two layers. The $42 billion senior layer is repaid first and carries less risk. A further $18 billion junior layer is led by Blackstone, which is taking $9 billion itself and selling the rest. Under the structure described in August reporting, Anthropic does not buy the chips directly. Investors pay for them and lease the hardware to the company.

From the prospectus to the marketBloomberg reported in August that Broadcom was in talks to raise more than $60 billion in debt, with Blackstone and Apollo discussing taking part. The three had done a similar $35 billion deal in June to expand Anthropic’s compute. Anthropic’s IPO prospectus then disclosed this week that Broadcom had agreed to lend it up to $42 billion. The new step is that the bank group is now actually looking for investors, and pricing and maturity have not been disclosed.

Broadcom's risk movesBroadcom used to simply sell chips. Now it is arranging its customers’ financing, and the August reporting said it would backstop part of the senior layer. If investors buy in smoothly, it locks in orders for years. If Anthropic’s revenue fails to keep up with its lease payments, the losses land first on these creditors and on Broadcom. How well this debt sells is itself a vote on demand for AI compute.

▪ SIGNALMore of the money behind AI compute now comes from bond investors than from tech companies, so whether chips get sold increasingly depends on whether Wall Street will hold the paper.

❯ Claude’s web app is rumored to have quietly switched to Fable 5.5, and Anthropic has not confirmed it

Same label, different answersSince October 1, a number of developers have said on social media that the model labeled Fable 5.1 in Claude’s web app and in Claude Code is behaving noticeably differently, and they suspect the backend now runs an unreleased Fable 5.5. Chinese outlets including 21st Century Business Herald picked up the story. Anthropic’s model directory still lists only Fable 5.1, and the company has not responded.

A one-line passwordThe most widely shared test is to ask the model whether it knows “Reset Guy Tibo” and tell it not to search. According to a rundown on SMZDM, Tibo is Thibault Sottiaux, who leads OpenAI’s coding agent Codex. When a Codex fault burned through users’ quotas, he announced a reset for paying users, and that became an in-joke. Users say Fable 5.1 through the API takes him for a French indie developer, while the web model gets it right.

The evidence is thinAccording to a roundup by Kingy AI, the posts show outputs such as 3D animations and voxel-style web pages, but nothing independently verifiable ties those requests to a new model. Neither a model’s self-description nor a trivia question proves that the weights changed. Anthropic released Opus 5.5 on September 22 and Sonnet 5.5 on September 28, which fuels expectations that Fable 5.5 is near, but any release date is still rumor.

What paying users care aboutIf the quiet rollout is real, consumers got a stronger model without knowing it, which is a gain. Enterprise customers, however, need to know exactly which version they are calling in order to reproduce results and estimate costs. A mismatch between label and reality, even in a test, unsettles teams that depend on stable output.

▪ SIGNALUsers now check a model’s identity by asking whether it knows last month’s memes, which shows how little the version label itself is trusted.

❯ Anthropic commits $100 million to a Claude Frontier Academy that aims to train 10,000 deployment engineers by the end of 2027

$100 million for 10,000 peopleAnthropic announced the Claude Frontier Academy on October 2, committing $100 million to train 10,000 “Frontier Deployed Engineers” by the end of 2027. These engineers work inside companies, wiring Claude into specific business processes and getting it to run in practice.

Trained like medical residentsThe program borrows from medical residency. Participants first attend several days of in-person training, run a simulated deployment with Anthropic engineers and pass an assessment. They then return to their own employers and spend 12 weeks leading a real Claude project before a final evaluation and certification. The first cohorts come from Accenture, Bain, Capgemini, Deloitte, McKinsey, Morgan Stanley, Commonwealth Bank of Australia and Novo Nordisk, among others, with sessions in San Francisco, New York and London.

Why run its own schoolAnthropic says the academy is meant to close the talent gap in deploying AI inside companies. More than 46,000 professionals had already earned certifications through its partner network. Consultancies that send staff can then use those credentials to deliver Claude projects for their own clients, which lets Anthropic extend its sales reach through other firms’ workforces.

The consultancies' calculationFor firms such as Accenture and Deloitte, certified engineers open the door to AI projects. It also ties their delivery capacity more closely to a single model vendor. If a client later wants a different model, how much of that skill carries over is a cost they have to weigh.

▪ SIGNALWhen a model company pays to train other firms’ engineers, the bottleneck in enterprise AI is no longer inside the model but in the stretch between the boardroom and the server room.

❯ Apple will tighten macOS Full Disk Access, saying AI agents make the permission riskier

Explicit action requiredAccording to TechCrunch, Apple says it will add new controls around macOS’s Full Disk Access permission, so that apps can obtain it only after “very explicit user action.” Apple says the risks tied to this level of access will grow substantially as AI agents become more capable and autonomous. It has not given a date for the new rules.

One permission, the whole computerFull Disk Access is the broadest permission in macOS. An app that has it can read a user’s files, mail, messages and browsing history. Apple says some developers are using it in ways that could expose everything on a system without users fully understanding what they agreed to.

Two cases in the backgroundAccording to The National, users had complained that Meta’s AI agent Muse read private messages without permission. OpenAI’s ChatGPT app for Mac launched a plug-in on August 20 that can read and send iMessages, which also requires this permission, and according to WIRED it patched a security flaw last month.

One more gate for agent makersAn AI agent that handles email and organizes files has to see that data. With Apple raising the bar for consent, installing desktop agents from Meta, OpenAI and others becomes more cumbersome, and some users will drop out at that step. Whether Apple’s own AI features face the same limits will determine whether this is a security measure or a platform advantage.

▪ SIGNALAI agents want to see everything, while operating-system security is built on granting only what is necessary, and one of the two will eventually have to give.

❯ Nvidia hits its first record high in more than four months as its market value reaches about $5.7 trillion

$237.88 intradayNvidia shares rose to a record $237.88 during trading on October 2, taking its market value to roughly $5.7 trillion, before closing at $233.95, up 1.3%. According to Yahoo Finance, it was the stock’s first record since May, and it is up more than 40% from its March 30 low of $165.17.

A buyback and a top pickOn September 28, Nvidia added $150 billion to its share buyback authorization, which the company called the largest authorization increase in history, ahead of Apple’s $110 billion in 2024, bringing the remaining authorization to $235 billion. On October 2, Morgan Stanley analyst Joseph Moore reinstated Nvidia as his top pick in semiconductors and kept a $300 price target.

Jobs data helped tooYahoo Finance attributes the rise to three things: the buyback; server maker Supermicro saying on September 23 that it had begun shipping new GPU racks; and Friday’s weak US jobs report, which showed 29,000 jobs added against 90,000 expected and cooled expectations of rate increases. Nvidia is the world’s most valuable company and last October became the first to pass $5 trillion.

Doubts at the topThe same day, The Information reported that AI-related borrowers have sold about $55 billion of high-yield bonds this year, but borrowing costs are rising and investors are getting choosier about data center projects. Nvidia’s orders ultimately depend on customers being able to borrow. If the bond market tightens and construction slows, the high valuation will be tested first.

▪ SIGNALNvidia’s record high and costlier data center financing arrived in the same week, meaning the stock market and the bond market are pricing the same AI buildout in two different ways.

❯ Nvidia launches a 64GB DGX Spark at $4,999 as the 128GB model climbs to $6,950

Half the memory, a higher priceNvidia has announced a DGX Spark with 64GB of unified memory starting at $4,999, on sale from October 23 through Acer, ASUS, Dell, Gigabyte, HP and MSI. According to PCMag, that is $1,000 more than the 128GB version cost when it launched last year at $3,999.

A small AI computer for the deskDGX Spark is a small desktop machine built on Nvidia’s GB10 chip that lets developers run and debug AI models locally without connecting to the cloud. The 64GB version uses the same chip, and Nvidia says it can run models of up to about 100 billion parameters. The 128GB version reaches about 200 billion.

Memory prices reach the end productAccording to Hardware Busters, the 128GB model has gone from $3,999 at launch to $4,699 early this year and now $6,950, a rise of about 74%, because of tight global memory supply. Data centers are buying up high-end memory and crowding out capacity for ordinary products, and Micron has said it expects the squeeze to last through 2028.

The cost of running models locallyDevelopers and small teams who want to run models on their own desks must now either pay the same money for half the memory or spend nearly $2,000 more. Memory directly sets how large a model the machine can run, so some will go back to paying for cloud usage.

▪ SIGNALThe memory that AI data centers are absorbing is making it more expensive to bring AI onto your own desk, because local and cloud computing are competing for the same chips.

❯ TSMC is reportedly exploring work with Musk’s Terafab and a possible presence in Texas

From complaint to cooperationAccording to Culpium, the independent outlet of former Bloomberg columnist Tim Culpan, citing people familiar with the matter, TSMC is exploring ways to work with Elon Musk’s Terafab project and help the new chipmaker run its factories in Texas. TSMC is also weighing a presence in Texas and could make Terafab the anchor customer of a new campus.

What Terafab isTerafab is a chip manufacturing project funded jointly by Tesla and SpaceX, sited in Grimes County, Texas. According to Yahoo Finance, the two companies filed an initial investment of $55 billion with the county in May, which could rise to $119 billion at full buildout. The chips are for Tesla’s self-driving system, Robotaxi and Optimus robots, and for SpaceX’s satellites and orbital data centers. Musk says he is building it because TSMC and Samsung are expanding too slowly.

Two options, nothing signedAccording to MacGeneration, relaying Culpium’s report, the companies are studying two options. In one, TSMC owns and operates the factory while Terafab takes a stake or guarantees purchases. In the other, Terafab keeps majority ownership and TSMC supplies technology and expertise. Nothing has been signed, and TSMC declined to comment. Musk had earlier indicated he wanted to use Intel’s 14A process for some chips. The full Culpium article is behind a paywall.

No shortcuts in chipmakingAccording to Startup Fortune, TSMC CEO C.C. Wei has said there are “no shortcuts” in the foundry business, and that a new fab typically takes two to three years to build and another one to two to ramp. If TSMC steps in, Terafab avoids some detours, while TSMC turns a potential rival into a customer. TSMC has already committed $265 billion to Arizona, so Texas would be an addition.

▪ SIGNALMusk built his own fab because TSMC was too slow and may end up asking TSMC for help, which says more about the barrier to advanced chipmaking than any earnings report.

❯ Robot “brain” maker FieldAI plans to raise $700 million, lifting its valuation from $2 billion to $10 billion in just over a year

Five times the valuationAccording to Business Insider, citing a person familiar with the matter, robotics software company FieldAI is raising $700 million at a $10 billion valuation, five times the $2 billion it was valued at in August 2025. The Next Web reports that a term sheet has been signed and the round has not closed.

Software only, no robotsFieldAI was founded in Irvine, California, in 2023. It develops what it calls a universal general-purpose robot brain: one set of software that can be installed in humanoids, robot dogs, drones and industrial rovers. According to TechCrunch last year, its models build physics into training so that robots adapt to new environments quickly and judge how confident they are and where the risks lie. The company makes no robot hardware itself.

Customers are already payingIts customers are construction firms, data center operators and defense companies. Reports say its revenue and signed contracts together exceed $135 million across more than 30 customers, up at least $35 million since June. Last August the company disclosed $405 million in total funding, including a round co-led by Bezos Expeditions, Prysm and Temasek.

Software valued above hardwarePhysical Intelligence is valued at about $11 billion and Skild AI at about $14 billion, both above most companies that build whole robots. Investors are betting that robot hardware will get cheaper and more alike, and that the scarce part is software able to fit many bodies. That bet needs more real deployments on sites and in factories to be proven.

▪ SIGNALCapital now prices robot software above robot makers, wagering that the industry will follow smartphones and send its profits to the operating system instead of the shell.

❯ MiniMax’s M3.1-Flash-Preview is used to reproduce Opus 5.5’s front-end showpieces, as flagship skills reach lightweight models faster

Reproduced within a weekAfter Anthropic released Claude Opus 5.5 on September 22, gacha games, hand-drawn animations and 3D scenes generated from a single prompt spread across social media. According to Sina Tech, people moved those prompts to MiniMax’s M3.1-Flash-Preview and got a ten-pull gacha game, a hand-drawn animation with a piano score and an interactive 3D spaceship interface, without repeated retries.

Only inside its own tools for nowM3.1-Flash-Preview went live on September 27 and 28. It is MiniMax’s lightweight model for everyday development, with a 1 million token context and text, image and video input. According to DataNorth, it is available only in MiniMax Code and subscription plans, and the company has published no pay-as-you-go price, speed figure or benchmark result.

Looking at images from step oneMiniMax’s explanation is that the M3 series learns images and text together from the very first step of pretraining, which required rebuilding its data pipeline and adding large amounts of interleaved image-text data. The model can therefore see what an interface should look like and write the matching code, instead of guessing through long chains of text reasoning. The report also notes it is still a preview and sometimes overdoes things.

Cheaper models take the routine workFor developers building web pages and small games, a lightweight model that produces a decent front end in one pass removes the need to call the most expensive flagship every time. The comparison so far rests on individual cases, not systematic evaluation, and there is no public data on how M3.1 Flash compares with Opus 5.5 on other tasks.

▪ SIGNALA flagship’s signature skill showed up in a lightweight model within a week, so the window in which a leader can charge a premium for something only it can do is now measured in days.

❯ TypeSafe says three-week-old Jev is used by about a quarter of the Fortune 500 and handles over a trillion tokens a day

Two numbers from the CEOTypeSafe CEO Diogo Almeida told The Wall Street Journal that the company’s model Jev is in use by about 25% of Fortune 500 companies, and that “we were at a trillion tokens per day about a week ago.” Jev was released only on September 15. The figures are the company’s own, have not been verified by a third party, and the definition of “in use” was not given.

Numbers, not wordsJev differs from familiar large language models. It does not generate text and returns only numbers. Asked a yes-or-no question, it gives a confidence score between 0 and 1. Asked a multiple-choice question, it gives a probability for each option. By its own description, it is mainly used for classification, routing and guardrails. It charges $0.042 per million input tokens and nothing for output.

Traffic that comes from machinesAlmeida calls the traffic “non-fleeting” because servers, not people, make the calls. TypeSafe was founded by Almeida, formerly of OpenAI, and two co-founders, and announced a $40 million seed round led by DCVC when it emerged on September 15. On October 1, Cloudflare released its open-weight Clef models with a Jev-compatible interface, so companies can swap them in without rewriting calling code.

A substitute within three weeksCompanies adopt Jev because much back-end judgment does not need a model that can write essays, and cheap and fast matter more. Switching costs for such models are also low, and Cloudflare’s compatible product means TypeSafe will soon compete directly on price and accuracy.

▪ SIGNALThe fastest-growing share of AI usage may not be people chatting at all, but machines making split-second judgments on behalf of other machines.

❯ The US Army sets up an autonomous systems command and gives unmanned equipment its own acquisition chief

One memoAccording to a memo obtained by Axios, acting Army Secretary Adam Telle signed an order on October 1 creating the Army Futures and Autonomous Systems Command (FASCOM) and designating an acquisition executive dedicated to autonomous systems, with priority on buying and fielding smart machinery.

Following Hegseth's projectsThe order came a day after Defense Secretary Pete Hegseth unveiled the Meridian and Agincourt projects on September 30. According to DroneXL, Agincourt is an interim effort to prepare a four-star Autonomous Warfare Command, targeted for October 1, 2027. Meridian is a 120-day study of future warfare co-directed by Elon Musk, Anduril co-founder Palmer Luckey and others.

Fielding by fiscal 2028According to Breaking Defense, the new executive must prioritize acquiring and fielding autonomous fires, combat vehicles, watercraft resupply and manned-unmanned teaming systems by fiscal 2028. The Army will also adjust force structure in six formation types, including aviation, armored, sustainment and training, and expand authorized positions for drone operators and robotics technicians.

Where the orders goA dedicated command and acquisition chief mean unmanned equipment no longer competes for budget scattered across programs, and defense technology companies that build autonomous systems face a clearer buyer, among them Anduril, which won a $20 billion Army counter-drone contract in March. The new command still needs legislation and appropriations from Congress, and the fiscal 2028 timetable depends on that.

▪ SIGNALOnce an army creates a command just for machines, the question about autonomous weapons has moved from whether to use them to who runs them and who buys them.

❯ An engineer uses GPT-6 Astra to break a Napoleonic-era cipher letter in six hours, and a 217-year-old puzzle is marked solved

An 1809 order read in six hoursCarter Church, an AI engineer at SentinelOne, used GPT-6 Astra to decipher, in about six hours starting from a single scanned image, a cipher letter written in March 1809 to French General Auguste de Marmont. CTech and Euronews reported the work, and he has published his transcription and scripts for others to verify.

155 signs, only 33 known beforeApart from one line of ordinary French, the letter consists of 24 rows of numbers, letters and invented symbols, about 1,300 cipher units and 155 distinct signs, of which researchers had previously identified only 33. It contains a military briefing ordered by Napoleon: the positions of French forces and Bavarian, Polish and other allied troops, written about two weeks before Austria attacked Bavaria in April.

How the model did itGPT-6 Astra first cut the scan into rows, read the handwritten signs and judged which were the same sign. It then used simulated annealing, a large-scale trial-and-error search, to match the remaining signs to letters and words, checking the output against French text by authors such as Dumas and Hugo. Church also reran the analysis with Napoleonic references excluded to confirm the model was not reciting. Satoshi Tomokiyo, who maintains the Cryptiana site of unsolved historical ciphers, has marked it solved.

A Vatican manuscript fell tooThis is not a one-off. A report in May said the Borg cipher, a 408-page manuscript in the Vatican library that had gone unread for more than 400 years, was solved with AI in 29 minutes by the Descrypt project, whose team includes Stockholm University professor Beáta Megyesi. It turned out to be a collection of old medical recipes. By contrast, a three-page cipher letter from Holy Roman Emperor Charles V had taken six months to break by hand. The project’s AI still relies on expert corrections to improve.

▪ SIGNALGeneral-purpose models have turned breaking old ciphers from a craft of a few specialists into something any patient person can attempt, so what history will run short of is verifiers, not solvers.

❯ After Trump renames AI “super intelligence”, Slovenia’s .si registrations jump from about 2,000 to 44,000 in September

More than twentyfold in a monthSlovenia’s domain registry, Registry SI, says about 44,000 new .si domains were registered in September, against about 2,000 in August, an increase of roughly 2,199%. According to the BBC and TechCrunch, 11,000 were added on September 30 alone, followed by nearly 13,000 in the next 24 hours.

A rush set off by an executive order.si is Slovenia’s country domain. On September 29, Trump signed an executive order requiring federal agencies to refer to “AI” and “artificial intelligence” as “SI” and “super intelligence”, and tech companies and speculators began grabbing addresses ending in .si. Domain seller Hostinger says .si briefly became the second most popular extension on its platform after .com, with more than half of buyers from the US and India.

Like .ai, but less lucrativeThe episode recalls .ai, which belongs to the Caribbean island of Anguilla and earned millions of dollars from the AI boom. But .si costs only about $12 a year at Hostinger, while .ai costs about $90 a year with a two-year minimum. Hostinger’s head of domains said the revenue goes to the registry, not the government, and that “Slovenia will not see what Anguilla saw.”

Mostly stockpilingReports say only about 3% of newly registered .si domains are explicitly related to artificial intelligence. Most of the rest were grabbed to be resold, and resales have already exceeded $1 million. Whether the new term will catch on outside government documents is uncertain, and the people paying now are betting that it will.

▪ SIGNALA term the industry has not yet accepted already has a domain business built around it, because speculators react to policy signals far faster than the industry itself.