AI · page 3

Anthropic IPO filing: $800bn compute bill, founders keep control

• Anthropic's S-1 sets up a Nasdaq listing under the ticker ANT as soon as next month. The FT says revenue has soared, though growth slowed in the third quarter, and a few big customers bring in a large share.
• It has signed compute deals worth more than $800bn, some running to 2045, with Broadcom, the big clouds, SpaceX, AMD and smaller providers. The founders call the approach bold but sane.
• A Long-Term Benefit Trust picks most of the board and the seven founders hold 50.1% of the vote. A Lex column says the past numbers tell investors little; ordinary holders should expect to be powerless.

Other side: A locked-in founder group and a mission trust can shield long-term bets from short-term traders, which some investors will welcome. But the filing itself warns growth may reflect early adoption; if demand cools, the compute bill does not.

See the KKReckons explainer →

Summary based on Financial Times. Read the full story ↗

Indian IT firms pitch tools to police rogue AI agents

• Companies that let AI agents book meetings, fix code and pitch clients now fear the bots could go rogue or reach data beyond their remit. TCS, Cognizant, HCLTech and Tech Mahindra all call managing and governing agents a new business line.
• TCS and Tech Mahindra are building software to coordinate agents and cap what they can do; HCLTech plans to buy Croatia's Robotiq.ai; Cognizant launched AI Nexus on 23 October, giving each agent its own ID and permissions and logging every task.
• HFS Research's Phil Fersht says the test is whether Indian firms can own a view on governance rather than resell a hyperscaler's stack.

Other side: Hyperscalers and AI labs build their own controls, so Indian firms risk becoming resellers again. Governance tools also tend to shrink into a feature once platforms mature, which would cap the fees.

See the KKReckons explainer →

Summary based on Mint. Read the full story ↗

Meta plans $95bn share sale, with Nvidia, AMD, Broadcom as anchors

• Meta plans about $35bn of underwritten share offerings and a further $60bn at-the-market programme, one of the biggest equity raises on record, to pay for AI data centres. The FT says the second part is expected to start in early 2027.
• Nvidia, AMD and Broadcom, which all sell it chips and gear, have shown non-binding interest in buying up to $5bn as cornerstone investors.
• A new share structure lets Meta sell stock without diluting Zuckerberg's control, which could meet shareholder pushback. The FT puts it in a wider rush to equity after Alphabet's raise and SpaceX's listing, with Anthropic next.

Other side: Selling shares is a safer cushion than piling on more debt, and Meta says AI already lifts engagement and ad returns. The open question is whether shareholders accept dilution while their say is held down.

See the KKReckons explainer →

Summary based on Financial Times. Read the full story ↗

Broadcom CEO rescued $63bn debt deal to lease chips to Anthropic

• A shell company will borrow $63bn to buy TPU chips Broadcom co-designs with Google and lease them to Anthropic, whose payments repay the loans. WSJ calls it the biggest AI infrastructure financing yet.
• As rates rose, some lenders wanted out. Hock Tan held weekend talks, agreed higher interest and firmer Broadcom guarantees to keep Bank of America, Citi and Morgan Stanley in; Blackstone funds part of a riskier, unguaranteed slice.
• Banks plan to sell half their share to bond investors only once the chips arrive next year, a risk if rates jump.

Other side: If Anthropic keeps paying, the guarantee never bites and lenders earn more for the trouble. But banks demanding higher pay to touch AI risk is a warning sign that matters more than the headline size.

See the KKReckons explainer →

Summary based on The Wall Street Journal. Read the full story ↗

Anthropic's S-1 pitches a $40tn market and warns of catastrophe

• WSJ reporters say the filing effectively sizes the prize at $40tn, the value of the world's knowledge work, and shows first-half revenue of $16.3bn. A huge net loss is mostly an accounting charge; the core business eked out a second-quarter operating profit.
• It warns models can try to avoid being shut down and that a catastrophe might be uninsurable. Goldman Sachs and Morgan Stanley lead the deal.
• The Journal's editorial board says the risk list reads like an effort to scare buyers, flags open-weight rivals and regulators, and suspects insiders want to cash out while the boom lasts.

Other side: The cash-out theory is the editorial board's guess, not something the filing says. Long risk lists are standard legal protection, and a lab sitting on a large cash cushion is hardly a forced seller.

See the KKReckons explainer →

Summary based on The Wall Street Journal. Read the full story ↗

Microsoft offers buyouts to up to 60,000 staff as AI costs climb

• Staff at vice-president level or below in some divisions in the US, Canada, UK and India can apply for severance, per a memo from people chief Amy Coleman. Xbox, Windows and Copilot are in; sales, security, data centres and quantum are out, a source told the FT.
• Unlike a July retirement offer to older staff, there is no bonus on top of standard severance, so take-up is expected to be low.
• Microsoft plans up to $175bn of AI infrastructure spending this year and free cash flow has fallen; a source says the offer reflects fast-changing work, not capex.

Other side: Voluntary exits keep goodwill and let people who already want out leave on their own terms. But the staff with the best options often take the money first, and Microsoft is still hiring in key areas.

See the KKReckons explainer →

Summary based on Financial Times. Read the full story ↗

India makes 20% of the world's data but hosts 3% of data centres

• India generates nearly 20% of the world's data but holds about 3% of global data-centre capacity, Mint reports, citing BCG. Google ($15bn, including a Visakhapatnam hub), Microsoft ($20.5bn) and AWS (over $21bn for 2026–2030) are moving to close the gap.
• AI racks draw 100–250 kW, against 5–10 kW for older servers, so new sites are built differently. Indian operators such as CtrlS and CapitaLand India Trust are adding capacity too.
• Mint argues hosting AI is not creating it: India's edge may lie in small local-language models for farming, health and finance.

Other side: Local compute cuts latency and cost for Indian startups, which is how ecosystems begin. And India may not need frontier-scale models to use AI well, as Mint itself argues.

See the KKReckons explainer →

Summary based on Mint. Read the full story ↗

Young US graduates fear AI is shutting them out of entry-level jobs

• In Pennsylvania's Lehigh Valley, the FT meets young graduates who say AI is closing the door to first jobs. A 2022 computer-science graduate says almost every interview now asks whether he is needed at all.
• Stanford research finds employment of software developers aged 22 to 25 down 19%, with customer service hit similarly, while jobs AI cannot do, such as health aides, are growing.
• There are no mass AI lay-offs yet; the squeeze is in hiring. One Lehigh student chose a master's degree over a job hunt to stay ahead of the machines.

Other side: Hiring freezes also follow rate rises and post-pandemic over-hiring, and the data show slower hiring, not firing. Earlier technology shifts also killed entry roles before creating new ones.

See the KKReckons explainer →

Summary based on Financial Times. Read the full story ↗

Who's paying for the AI build-out? Savers, pensions and insurers

• FT columnist Robert Armstrong argues that almost every big pool of capital in the rich world is now funding AI data centres, so anyone with savings is exposed.
• He cites a Brookings estimate that the US build-out will cost $10.3tn through 2032, more each year than the railroad and interstate booms combined. Hyperscalers are spending all their cash flow, chipmakers are buying stakes in AI labs, and debt is piling up.
• Most of that debt is rated investment grade, thanks to hyperscaler guarantees or chips as collateral, which lets it into pensions, annuities and life insurance.

Other side: Investment-grade ratings reflect borrowers that are among the most profitable firms on earth, and booms that looked reckless, like the railroads, still left valuable assets behind. The risk is timing, not ruin.

See the KKReckons explainer →

Summary based on Financial Times. Read the full story ↗

Why US shoppers suspect retailers use AI to raise prices

• American shoppers increasingly believe stores, restaurants and ride-hailing apps use AI and personal data to charge them more, the WSJ reports. One Walmart app user saw a hard drive listed $110 cheaper on his wife's phone; Walmart blamed a since-fixed error.
• Walmart, Kroger and McDonald's deny pricing by who is buying. Walmart's chief executive wrote: "We price the product, not the person."
• Lawmakers are moving anyway. New York requires disclosure of algorithmic pricing, Maryland, Connecticut and New Jersey restrict data-driven pricing, and the FTC has proposed penalising deceptive personalised prices.

Other side: Studies of digital shelf labels in the US and Europe find they mostly enable markdowns on food near expiry, not surge pricing. Colorado's governor vetoed a ban because it would also have outlawed targeted discounts.

See the KKReckons explainer →

Summary based on The Wall Street Journal. Read the full story ↗

CoreWeave's debt pile tests whether the AI build-out can be financed

• CoreWeave, an AI cloud firm built on borrowed money, has become a test of whether the AI build-out can be financed. Investors have been selling its shares and bonds.
• Its model: sign a contract with a big customer such as Microsoft, then borrow against it, with chips as collateral. Other borrowing is repaid from corporate cash, so debt sits on top of debt; the total is $35.4 billion.
• Rising Treasury yields and hot inflation raised borrowing costs, and yields on its existing debt passed 15%. One startup CEO says Goldman bankers could not explain the slide.

Other side: CoreWeave's CEO calls it a market misunderstanding, Nvidia's Jensen Huang voices confidence, and Goldman's analysts upgraded the stock. Project loans are repaid from customer contracts, so a falling share price is not the same as default.

See the KKReckons explainer →

Summary based on The Information. Read the full story ↗

SpaceX-Tesla merger and Anthropic IPO: how investors price AI

• The FT's Big Read says Elon Musk wants SpaceX and Tesla combined into one AI-led company worth about $3.5tn, with pay awards that could reach $1.7tn.
• Anthropic told investors its third-quarter revenue was $16.8bn, up 45% on the quarter, and profitable on an adjusted basis, with an annual run rate of $67bn, Damodaran notes.
• Damodaran says young, loss-making firms are priced on buzzy metrics, from eyeballs to users to ARR, and warns that ARR can be defined in different ways.

Other side: Bulls say AI's market is vast and the leaders will take most of it. Bears say compute costs, cheaper open-source rivals and rules could squeeze margins, and critics see the merger as a way to shift attention from a weaker Tesla. Damodaran takes no side until the full prospectus is filed.

See the KKReckons explainer →

Summary based on Financial Times, The Economic Times and Musings on Markets. Read the full story ↗

TCS and Infosys: AI tops 10% of revenue, yet pricing bites

• TCS and Infosys both say AI brought in more than 10% of revenue in Q2 FY27, and most of India's top five IT firms widened operating margins.
• Clients want the savings passed on. Infosys cut its FY27 growth guidance to 1-2% in constant currency, from 1.5-3%, and says it will favour profitable deals over growth.
• CEO Salil Parekh sees a big cyber opening as AI models uncover flaws, and says Infosys is already sharing defences with clients, though budgets may not rise much.

Other side: Less effort need not mean less profit, as HFS Research's Phil Fersht argues, if firms price on outcomes. And clients may fund cyber work by trimming elsewhere.

See the KKReckons explainer →

Summary based on Moneycontrol. Read the full story ↗

AI finds bugs faster and attackers follow in 'Black October'

• The WSJ reports a wave of attacks using AI agents and hacking tools, adopted by China- and Russia-linked groups and ordinary criminals.
• Reuters spoke to seven cybersecurity professionals who say AI-found flaws will almost certainly cause more outages like this week's PHP-related ones on government sites. VulnCheck says flaws found doubled year on year, with AI a driver.
• Models can scan code round the clock and turn exploits around in hours. Experts split on whether attackers gain: curl's lead developer says AI-found bugs tend to be exotic.

Other side: More disclosed flaws do not automatically mean more successful attacks, and AI also helps defenders. A Hunt.io researcher said the criminals he tracked had not yet built a working PHP exploit by Thursday afternoon.

See the KKReckons explainer →

Summary based on The Wall Street Journal and Reuters. Read the full story ↗

PHP2Shell flaws expose the old code behind 7 in 10 websites

• Two flaws called PHP2Shell, disclosed on Wednesday, sit in the part of PHP that handles crashes. PHP runs WordPress and about 7 in 10 websites, and the bug opens a server before a site's own code or firewall runs.
• AI found them: security firm Calif, using an OpenAI model, caught the first in August, and Wiz found both two weeks later. Cloudflare saw millions of attack attempts by the night of the announcement.
• Patches exist for PHP 8.2 to 8.5, but about 7.5 lakh of roughly 5 million PHP servers run versions with none. Finland, Wikipedia and Japan's tax agency took services offline. Pravin Kaushal argues the lasting fix is funding PHP's few maintainers.

Other side: The same AI tools that found PHP2Shell can help defenders audit and patch faster. Funding maintainers helps, but site owners who never update stay exposed however well PHP's team is paid.

See the KKReckons explainer →

Summary based on The Economic Times. Read the full story ↗

Tech Mahindra ties up with SpaceX's AI arm to modernise old software

• Tech Mahindra, India's fifth-largest IT firm, has partnered with SpaceXSI, the AI division of SpaceX, to help companies modernise legacy applications and speed up software development.
• It will also use the Grok Build chatbot to create industry-specific tools, starting with technology, telecom, and banking and financial services clients.
• Tech Mahindra already has similar pacts with Google Gemini and Microsoft; Reuters says this is the first time an Indian IT major has teamed up with SpaceX's AI unit.

Other side: Partnership announcements are cheap, and Tech Mahindra already has similar deals with Google and Microsoft. Revenue, not logos, will show whether clients pay for AI-led modernisation.

See the KKReckons explainer →

Summary based on Reuters. Read the full story ↗

Hedge funds post broadest losses since Covid as AI stocks slide

• Bloomberg reports, citing Goldman Sachs's prime brokerage, that Thursday's tech sell-off pushed all three broad equity hedge fund strategies down more than 1% on the same day for the first time since Covid; about two-thirds of funds lost money.
• In FT's Unhedged, Robert Armstrong says Treasuries rallied hard, with the 10-year yield falling to 4.96%, and calls it a fear-driven, risk-off move rather than a bet on lower inflation or rate cuts.
• He notes losses are concentrated in hyperscalers, chipmakers and data-centre builders, while defensives such as consumer staples have rallied; strong consumer spending and earnings still support stocks.

Other side: Armstrong notes the economy and earnings remain strong, and most S&P 500 stocks are up since early October. One bad day in earnings season may be a rotation, not the start of winter.

See the KKReckons explainer →

Summary based on Bloomberg and Financial Times. Read the full story ↗

Meta's Muse passes 10 million downloads, but costs outrun revenue

• Zuckerberg said on Meta's earnings call that more than 10 million people in the US and Canada have downloaded Muse, its AI agent app, far faster than ChatGPT's app reached that mark.
• WSJ columnist Asa Fitch argues that even rapid subscriber growth would cover only a slice of Meta's AI spending, which analysts expect to approach $200 billion next year; free cash flow has turned negative and debt has climbed.
• Meta also created a new class of nonvoting shares, seen as a step towards raising equity, and its stock fell sharply on Thursday. Fitch adds that low trust in Meta could limit what users let an agent do.

Other side: Meta's ad business still grew strongly last quarter, and its reach across Instagram and WhatsApp is unmatched. Heavy early spending is how platforms are won; judging returns six weeks in may be premature.

See the KKReckons explainer →

Summary based on The Wall Street Journal. Read the full story ↗

Microsoft's AI beats doctors on hard cases; AI phage given to patient

• Reuters reports that a Microsoft AI tool that works up patients step by step, choosing questions and tests as it goes, was more than twice as accurate as general physicians on over 300 complex cases in a study in Science, and its work-ups cost less.
• The Independent reports that California lab Radical Numerics says it gave phages designed by its Omnii-1 model to a patient with an antibiotic-resistant infection under compassionate use; outcomes are undisclosed and await peer review.
• Microsoft will next test emergency-room triage with Beth Israel Deaconess in Boston.

Other side: Benchmark cases are not hospital wards, and the doctors in the study worked with textbooks and the internet rather than colleagues. The phage claim has no published outcome. Trials and regulators will decide.

See the KKReckons explainer →

Summary based on Reuters and The Independent. Read the full story ↗

Big US banks curb AI agents after unauthorised account access

• Axios reports that big US banks are finding third-party AI agents inside customer accounts, usually logged in with the customer’s own password, and some acting beyond instructions. Morgan Stanley is blocking them; JPMorgan, Bank of America and Citi are restricting them; most banks lean towards read-only access.
• In an FT column, John Thornhill argues firms must ‘Know Your Agent’: identify it, keep it behaving as intended and settle who is liable when it errs. He cites Deloitte: 74% of companies plan to use agents within two years.
• 404 Media shows ID checks already strain: a scammer opened a CashApp account in Amanda Smith’s name, and she was charged.

Other side: Customers will want agents to pay bills and move money, and bankers admit a middle ground is needed. Blanket bans may only push password-sharing out of sight; vetted, read-only or signed agent access could prove safer.

See the KKReckons explainer →

Summary based on Axios and Financial Times. Read the full story ↗

AI researchers leave OpenAI, DeepMind to build startups in India

• Inc42 tracks researchers leaving global labs to build in India: Shyamal Anadkat, who led OpenAI’s Applied Evals team, founded Empirical Machines for AI-driven molecule and materials research; ex-DeepMind engineer Parth Sarthi’s Chariot builds speech-native models under the IndiaAI Mission.
• Former Google engineering director Anish Das Sarma’s Reinforce Labs red-teams enterprise AI. Empirical Machines is reportedly in talks to raise about $15 Mn, per Moneycontrol.
• Investors say lab pedigree helps recruit talent but does not prove a business; DeVC’s Rahul Mathur asks who will pay for the research.

Other side: Pedigree is not product. Research takes years to pay off, compute is costly, and investors say customers, not résumés, will decide which of these startups survive.

See the KKReckons explainer →

Summary based on Inc42. Read the full story ↗

AI bull market is no dotcom bubble, argues Bloomberg’s Nir Kaissar

• In a Bloomberg Opinion column carried by Mint, Nir Kaissar argues the AI rally lacks the euphoria that drives bubbles: bears now outnumber bulls by 8 percentage points in AAII surveys, while bulls led by over 20 points in the late 1990s.
• He concedes AI’s grip is bigger: AI firms make up 40% to 60% of US market value and nearly 70% of S&P 500 third-quarter earnings growth, per Citadel Securities estimates.
• His forecast: profit growth cools and a pullback follows, but absent a crisis, nothing worse than an ordinary bear market.

Other side: Scepticism alone does not prevent a bust. Concentration is higher than in 1999, earnings can disappoint, and as Kaissar admits, bubbles are hard to call in real time.

See the KKReckons explainer →

Summary based on Mint. Read the full story ↗

Terence Tao calls for ‘Maths 2.0’ as AI floods the field with proofs

• In a Nature World View column, Fields medallist Terence Tao says AI-aided proofs have appeared for hundreds of long-standing problems in recent months, and calls for a ‘Maths 2.0’ no longer defined by scarce proofs.
• He flags costs: OpenAI’s 6 October release of hundreds of solutions disrupted research programmes; many AI results miss publication standards; credit for prior work is unclear; juniors lose a route to recognition.
• He warns understanding lags output, ‘proof indigestion’, and sees similar risks in opaque models used for bank lending.

Other side: Tao uses AI in his own research and says that, used creatively and with the right incentives, it could deepen understanding. The upheaval may be a transition cost, not a lasting loss.

See the KKReckons explainer →

Summary based on Nature. Read the full story ↗

PHP2Shell, eMudhra, PingCAP: one week, three security scares

• PHP2Shell: two critical flaws (severity 9.2 of 10) let attackers take over unpatched PHP servers without a password. Found with AI help, fixes shipped on 21 Oct, a day early after a leak; Amazon says they are now broadly exploited, and the NSA says patch or go offline.
• eMudhra: eight intermediate certificates had keys weak enough to recover almost instantly. Mozilla, Google, Apple and Microsoft have blocked them; no misuse shows in public logs, and India’s IT ministry has set up a probe.
• PingCAP: stolen service credentials let an intruder copy TiDB Cloud user details, SQL editor text, billing data and encrypted API keys. It sees no sign that data stored in TiDB was hit.

Watch next: Patch speed. Servers and apps that update slowly may show damage in the coming weeks, though so far no attack using the eMudhra keys is known.

See the KKReckons explainer →

Summary based on SecurityWeek and PingCAP. Read the full story ↗

Want to short the AI boom? Futures on chip rental prices are coming

• Silicon Data, founded by former Bloomberg data executive Carmen Li, compiles indexes of Nvidia chip rental prices. CME plans cash-settled futures on them for H100 and B200 chips, running up to three years out.
• The launch slipped after the CFTC extended its review; the regulator has flagged manipulation risk and doubts that compute is standard enough to trade like a commodity.
• Rival Ornn, backed by a16z, is working with ICE on similar contracts. Silicon Data’s backers include DRW founder Don Wilson, VanEck and Jump Trading.

Other side: Sceptics recall dot-com bandwidth futures, which failed when a fibre glut sank prices; a market only works if its benchmark can be trusted.

See the KKReckons explainer →

Summary based on The Wall Street Journal. Read the full story ↗

OpenEvidence: the $15bn AI start-up reshaping how US doctors treat

• OpenEvidence, founded by Daniel Nadler, runs an ad-funded chatbot that searches peer-reviewed journals and clinical guidelines to help doctors diagnose and treat patients.
• Nadler says nearly two-thirds of US physicians use it. It raised money in September at a $15bn valuation, and hospitals including NYU Langone and Mount Sinai have built it into their record systems.
• Evidence on accuracy is mixed: a Mayo Clinic study (Mayo is an investor) found accurate advice, while a June Nature paper found it trailed general models such as Gemini and GPT on clinical benchmarks.

Other side: Doctors adopted it on their own, which suggests it saves real time; the worry is juniors leaning on it before mastering the basics, and OpenAI and Anthropic are moving in.

See the KKReckons explainer →

Summary based on Financial Times. Read the full story ↗

Anthropic's IPO filing: hot growth, a huge loss and a narrowing window

• WSJ reports Anthropic's draft prospectus shows a first-half net loss of $80.7bn, mostly an accounting charge tied to the value of its convertible notes rather than cash spent running the business.
• Revenue is growing far faster than computing costs, but a big share comes from a handful of customers, and the firm signals it could spend up to $750bn on compute, partly through deals still being negotiated.
• An FT Unhedged column argues the listing window, now pencilled in for November, is narrowing as markets wobble, borrowing costs rise and questions grow over who is liable when AI agents cause harm.

Watch next: Most of the loss is a non-cash accounting charge and revenue still outpaces costs, so the headline may overstate the strain. The sharper test is reliance on a few customers.

See the KKReckons explainer →

Summary based on The Wall Street Journal and Financial Times. Read the full story ↗

CoreWeave's bond yield and CDS: an early warning for AI debt?

• Some CoreWeave bonds fell sharply in mid-October, pushing the yield on its 2031 notes above 16%. The coupon is unchanged, but new buyers now demand a far bigger return for the risk.
• The cost of five-year credit default swaps, insurance that pays out if a borrower fails to pay, hit a record 1,194.5 basis points on 15 October.
• ET Prime explains that rises in both signals show confidence weakening. Before the 2008 crisis, rising bank CDS gave early warning; the worry now is stress spreading to AI lenders, suppliers and stocks.

Watch next: CoreWeave's November results. Markets often reprice risky debt without a crisis, the firm reports a large contracted backlog, and no one has forecast a default.

See the KKReckons explainer →

Summary based on The Economic Times. Read the full story ↗

Microsoft's exit from an Nscale lease is a reality check for neoclouds

• Microsoft pulled out of leasing capacity at Nscale's planned data centre in Essex, which may not get power until after 2030, as Nscale targets a New York listing at a $35bn valuation.
• An FT Lex column says neoclouds need power, chips and upfront financing at once, so their big order books carry heavy execution risk; Anthropic and Microsoft fill most of Nscale's backlog.
• Valued on peers' multiples, Lex argues, Nscale is worth far less, while bigger players with ready capacity stand to win.

Other side: The Essex site was a small slice of Microsoft's spending with Nscale, and its other contracts may hold; one cancelled lease need not sink the model.

See the KKReckons explainer →

Summary based on Financial Times. Read the full story ↗

Ares and Gemes: AI debt looks safer than it is

• Ares’ Fall 2026 ‘In The Gaps’ says insurers, long the market’s most cautious investors, now hold venture-style risk packaged as investment-grade bonds funding the AI build-out.
• Ares traced 100+ financings back to eight names, from Meta and Oracle to OpenAI and Anthropic, and notes chip resale values are assumed anywhere from 10% to 60%.
• Orlando Gemes’ LinkedIn summary adds early warnings: Meta-linked Beignet bonds hit a record low near 94.4, and Oracle sent a force majeure notice on its Project Jupiter campus.

Other side: Gemes says none of this is a default yet, Oracle calls its notice precautionary and Blue Owl says its commitments are unchanged. Is this a structural flaw, or partly just rates?

See the KKReckons explainer →

Summary based on LinkedIn and Ares. Read the full story ↗