The Read

Apollo's chief economist put a number on the AI payoff gap, and the number is zero. Torsten Slok's five-chart note shows real revenue per employee at the S&P 493 flat at roughly $550,000 since 2019, while the Magnificent Seven climbed from about $550,000 in 2022 to nearly $800,000 — and margins rising for those seven since the third quarter of 2023 and flat for everybody else. Read as a verdict on the technology, that is bearish. Read as a control group, it is the day-zero thesis with a denominator: seven companies re-founded their work around cheap intelligence and moved revenue per head about 45% in four years, while 493 bought seats, bolted it onto the existing workflow, and got nothing a chart can see. Two other items landed in the same twenty-four hours and belong to the same story — insurers now writing five-million-dollar-a-day delay cover for gigawatt data centres, and Japan's biggest domain and hosting provider asking its government for active cyberdefence. That is what a build-out looks like when it stops being an experiment and becomes plumbing: boring institutions arrive, and they arrive with prices attached.

Tide

One confirmation. ai-as-worker gets its control group, and the control group is the argument for re-founding rather than against the technology. No tide shifted.

ai-as-worker, confirmed by its own control group: the seven that re-founded moved revenue per head about 45%, the 493 that bolted it on moved zero

Torsten Slok, chief economist at Apollo Global Management, published a five-chart note arguing that AI "has delivered zero financial payoff to corporate America outside of tech." His measures, inflation-adjusted to the fourth quarter of 2019: revenue per employee at the Magnificent Seven — Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia and Tesla — rose from about $550,000 in 2022 to almost $800,000 in 2026, while the remaining S&P 493 has been flat at around $550,000 since 2019. Profit margins have risen for the seven since the third quarter of 2023 and stayed flat for the rest. On the San Francisco Fed's utilisation-adjusted total-factor-productivity series, which he reproduces, he says there are "still no signs" of AI lifting productivity growth, and he argues mainstream firms are starting to treat token spend as a cost centre rather than a growth input. His own hedge is worth keeping: turning cheaper intelligence into growth "may be harder than it sounds," which is a claim about difficulty, not impossibility. Two pieces of context cut against the strong version. The Federal Reserve Bank of New York published work by Jeremy Pearce on Thursday 15 October finding real revenue per worker at US public firms grew 2.2% a year from 2019 to 2025 against 0.7% a year from 2010 to 2019, with operating margins up from about 16% to about 18% of sales — the same headline metric on a different sample and a different window, since Pearce pools all US public firms without separating the seven from the 493, so the two results can both be true. And Slok's spend charts rest on Ramp's AI index of 14 October, where adoption reached 55.9% in September but token spend showed virtually no growth in August or September and the top 1% of spenders averaged $9,568 per employee per month, down 3.1% from a July high of $9,873 — which Ramp itself attributes to frontier price cuts, not retreat. Ramp also reports that 1% of firms drive 80% of OpenAI and Anthropic spend, so this is a concentrated sample, not corporate America.

So what: Stop arguing about whether AI works and start reading this as a map of where the returns sit. The gap between $550,000 and $800,000 per employee is not a gap in model access — every one of the 493 can buy the same tokens at the same published prices, and the prices keep falling. It is a gap in whether the work was rebuilt or merely assisted. The operator test for any AI line item in your 2027 budget: name the process that will no longer exist when this ships. If the answer is "the same process, faster," Slok's flat line is your forecast. If the answer is a function that gets re-founded, you are in the other cohort, and the cost of joining it has never been lower.

AI delivers 'zero financial payoff' to most companies, Apollo economist says
AI Provides 'Zero Payoff' to S&P 493 Outside Tech, Apollo Economist Says
Apollo Economist Finds Zero AI Payoff Outside Big Tech
Revenue per Worker Is Up. Are Firms Getting a Larger Share?
October 2026 Ramp AI Index: Cracks in the AI thesis, part 3

Waves

Late delivery on a data centre now costs $5 million a day, and the insurance market quintupled its own product in a matter of months

The Information's Ann Davis Vaughan reported on Sunday from inside the brokers writing cover for the build-out. Mike Mathews, who leads global digital infrastructure at Marsh, designed a bespoke "delay-in-delivery" policy this spring to protect a developer from a $1-million-a-day late fee — then watched tenants in new deals start asking for upwards of $5 million for each day of delay. "Sent a shockwave," he says. "I'm not saying investors were hands off on insurance, but they're all hands on now." The scale around it: S&P Global estimates data centre insurance demand could generate about $10 billion in new premiums in 2026 — roughly twice the global aviation insurance market's own annual premium of around $5 billion — and Swiss Re estimates total data centre premiums could approach $24 billion by 2030. Construction is at roughly $2 trillion of insurable value growing $300 billion a year, and S&P warns that it expects "capacity constraints to limit the insurance industry's ability to fully insure hyperscale data center projects," with total insurable values reaching $20 billion to $30 billion per location — against facilities, chips and servers costing roughly $50 to $60 billion per gigawatt. The arithmetic of a slip: on a one-gigawatt shell delivered two months late at $5 million a day, the developer owes about $300 million in penalties while forgoing rent, which is why KKR's global head of digital infrastructure, Waldemar Szlezak, wrote this month that "late penalties will bankrupt more developers than interest rates ever will." Operational exposure is being priced too — RBC Capital Markets recently reported a data centre risking a $150 million penalty after temperatures deviated from safe targets for a month, and Gallagher Re found a single 45-minute outage could wipe out 40% of annual cash flow against a 99.999% uptime promise. Marsh and Aon are both pooling risk across projects because no single insurer can absorb tens of billions on one campus, and S&P expects growth in insurance-linked securities as a result. Mathews says he has not yet found a project to be uninsurable, and reinsurers have excess capital they want to put to work.

So what: This is the compute-financialization wave hitting the layer where risk gets priced by people who have to be right, and that is a maturation signal rather than a warning. Insurers pricing delay, GPU residual value and utility-interconnect failure means the build-out is legible enough to underwrite — which is precisely what lets pension money and bondholders fund it instead of only balance sheets. Two things follow for anyone signing compute contracts. First, read the delay clause before the price clause: a $5-million-a-day penalty is now a normal ask, and whichever side absorbs it is carrying the real economics of the deal. Second, if you are a tenant, the late fee you negotiate is an insurable asset for your counterparty, so ask who is carrying it and whether the policy actually exists — Marsh's own team says GPU residual-value cover is still nascent.

Data Center Late Fees Hit $5 Million or More Per Day, Spawning New Insurance
Data center tenants now demand $5M+ a day in late-delivery fees, up from $1M this spring, spawning new "delay-in-delivery" coverage from brokers like Marsh
Insurance capacity constraints for data centres to drive ILS use: S&P

Japan's biggest domain and hosting provider asked its government to go on offence, and is paying every customer ¥5,000 for the trouble

Reuters reported on Sunday that GMO Internet will compensate customers for losses from unauthorised access to its foreign exchange trading service, with every customer receiving ¥5,000 — about $32 — as what CEO Masatoshi Kumagai called in a blog post "an initial gesture of good faith." GMO also asked the Japanese government to apply cyber defence measures in response to the attack — the Japan Times and Nikkei Asia, reporting it over the same weekend, name the specific policy GMO wants, Japan's "active cyber defence" doctrine. Reuters further disclosed that GMO received a demand for money with a threat to publish company and customer data, and said it will not pay and is working with police. The market took it seriously as soon as it could: against a $2.6 billion market value at Friday's close, GMO shares fell 11% in Tokyo morning trading on the next open, with the benchmark index flat. It is not alone — Park24, which runs the Times car-sharing service, said on Friday it would pay customers ¥3,000 after a personal data breach, and on the same Friday Japan said it would stand up a team to simulate cyberattacks using AI to find weaknesses in its own technology systems. Reuters ties the wave to the PHP flaws that drove millions of hacking attempts and took government and company websites offline in Japan and Europe the previous week, and quotes experts saying AI automates the scanning and phishing work that makes this faster, cheaper and harder to detect. Worth holding alongside the honest counter-read, which gets its own ripple below: on Interconnects' Sunday podcast, Nathan Lambert argued the vulnerability-discovery and time-to-exploitation curves were already exponential before any of this, and that mass exploitation "hasn't happened and isn't happening yet."

So what: The ai-security-refounding wave just produced its clearest institutional artefact: a listed infrastructure company publicly asking the state for offensive capability, and a government answering with an AI red team of its own. That is the shape of the resolution, not the shape of the collapse — the same capability that found the holes is what closes them, and it is now being funded at state scale. Two moves. If you sell infrastructure, write the compensation policy before you need it; GMO paid a flat goodwill amount within days and took an 11% hit, and the firms that improvise this under pressure will pay more for less credit. If you buy it, the diligence question has changed from "are you certified" to "when did you last patch, and who scanned you with what."

Japan companies compensate users as cyberattacks buffet economy
GMO asks government to take active cyberdefense measures
GMO Internet CEO Masatoshi Kumagai apologizes for cloud security breach

The leaked prospectus becomes a filed one: Anthropic's S-1 could go public as early as Monday, and the open questions left are structural rather than financial

The Information's Nick Wingfield wrote on Sunday that Anthropic "could make public the paperwork for its initial public offering as early as Monday" — reported, and not confirmed by the company. Because the Wall Street Journal already viewed a preliminary version and the numbers were reported out on Tuesday 20 October, Wingfield's point is that much of the financial reveal has already leaked; what the filing would newly settle is structure. His colleagues' shortlist: cash flow and the trajectory to profitability; how much Anthropic actually pays Google and Amazon to serve its models, and what those companies' stakes are now worth; revenue concentration, where the already-reported fifth-from-two-customers figure needs a more recent number, and the split between API, Claude subscriptions and resale through partners; the mechanics of a governance structure in which the non-shareholder Long-Term Benefit Trust appoints most of the board while the seven co-founders have a plan to hold a combined 50.1% of voting power even after selling down, reportedly with a special employee share class for tiebreakers that public reporting on Anthropic's share structure has not yet described; and whether lockups run past the usual 180 days, with a possible secondary-sale exception at listing that this year's SpaceX and Cerebras IPOs did not permit. The same note flags the week's other scheduled tests: Google and Microsoft report on Wednesday 28 October, Amazon on Thursday 29 October.

So what: The interesting disclosure here is not the loss line, which is already public and mostly non-cash. It is whether a company can carry a mission-control structure into public markets and have index money buy it anyway. If the trust and the founders' 50.1% survive the filing intact and the book still clears, that is a template every AI company with a charter will copy, and it changes what founders can keep. Practical read for anyone with Anthropic in their stack: the supplier-concentration and compute-obligation disclosures are the ones to read, because they tell you how much of your vendor's cost base is contractually fixed for a decade — which is the real determinant of whether the price cuts you are currently enjoying are a strategy or a subsidy.

Anthropic's Big Reveal
Anthropic revenue hits $16.3 billion as preliminary S-1 outlines IPO risks - WSJ

Britain has three departments for data centres and none of them owns the power, and a minister told Parliament that is clear

The Register's Lindsay Clark reported on Sunday morning on evidence given to the House of Commons Science and Technology Committee this week by Liz Lloyd, parliamentary under-secretary of state. After the abolition of the Department for Science, Innovation and Technology in Prime Minister Andy Burnham's reshuffle, industry liaison on data centre construction sits with the Department for Business, Innovation, Science and Trade; telecoms and digital infrastructure with the Department for Digital, Culture, Media and Sport; power supply, by The Register's reckoning, with the Department for Energy and Net Zero; planning permission with local government; and overall planning policy with the Ministry of Housing, Communities and Local Government. Lloyd's assessment of the arrangement in sector terms: "that is clear." The practical consequence arrived earlier this month, when Microsoft pulled out of plans for Britain's largest AI supercomputer because it could not secure power. Asked about the government's £45 billion productivity figure — first published in a January 2025 press release headed "Archaic tech sees public sector miss £45 billion annual savings" — Lloyd said it was "an estimate of long-term productivity gains," not a target and not a cashable saving.

So what: Set against the Marsh item above, this is the whole datacenter-backlash wave in one page: private capital has priced interconnection delay at $5 million a day, and the public counterparty that controls the interconnection is split across three ministries with no single owner. That asymmetry is the actual bottleneck in the UK, not planning sentiment and not capital. It is also the clearest opening on the board — the jurisdictions that consolidate siting, grid connection and planning into one desk will win the next decade of this build-out at the expense of the ones that do not, and several already are. If you are picking a European site, make "name the single official who can commit to an energisation date" a gating question, and treat the absence of an answer as a priced risk rather than a procedural annoyance.

Want to build a UK datacenter? Please hold while Whitehall transfers you

Ripples

A recruitment company's lab shipped the best open agentic coder in the 30B class, Apache 2.0, and used a jobs benchmark to prove it

Nanbeige LLM Lab, which sits inside the Chinese recruitment platform BOSS Zhipin, released Nanbeige4.3-33B with open weights on Hugging Face over the weekend, with MarkTechPost writing it up on Sunday. It is a 33-billion-parameter dense model — no mixture-of-experts — under Apache 2.0, with a 524,288-token context window, positioned on its own model card as a "flagship teacher model" meant to anchor distillation into smaller models such as a planned Nanbeige4.3-3B. The card scopes its own claim to "state-of-the-art agentic performance among open models in the ~30B parameter class," and that scope is the right one. The vendor-reported numbers, measured against Qwen3.8-27B, Meta's Muse-Glimmer-30B and Google's Gemma4-31B at highest thinking effort: Terminal-Bench 2.1 at 80.9 against 73.0, 61.8 and 47.2; Terminal-Bench 4.0 at 23.8 against 6.3, 1.6 and 0.0; DeepSWE at 65.5 against 42.5; Automation Bench at 54.2 against 38.3; JobBench at 44.8 against 34.5. On reasoning the margins collapse: AIME 2026 at 95.2 against Qwen's 94.5 and Muse-Glimmer's 94.7, which is inside the noise. And it loses three outright — SWE-Bench Pro (60.3 to Qwen's 61.7), GDPval (84.8 to 86.4) and GPQA Diamond (89.5 to Gemma4-31B's 91.3 in Nanbeige's own table). Read all of these as vendor benchmarks run on the vendor's harness, and note that MarkTechPost flags some competitor rows as differing from those labs' own published cards. The one that is hard to look away from is YC-bench, a simulated-earnings measure, where Nanbeige reports $1,025,437 against $641,290 for Qwen3.8-27B.

So what: Download it and run it against your own agentic coding harness this week, because the price of the capability is now zero and the licence is permissive. The strategic point is who shipped it: a hiring platform built the best open ~30B agent model because it had the one asset the frontier labs do not — a corpus of what work actually looks like across occupations, which is why JobBench and a 35-occupation breakdown are the benchmarks it chose to lead with. That is the day-zero pattern in its purest form, and the lesson generalises: proprietary process data about a specific domain of work is a model-training asset, and most companies sitting on it have not noticed.

Nanbeige Releases Nanbeige4.3-33B, a 33B Open Agentic Coding Model
Nanbeige/Nanbeige4.3-33B

Hassabis has cut his AGI estimate twice this year, from five-to-ten years to two-to-three

Azeem Azhar laid out the ladder in Exponential View's Sunday issue: at Davos in January, Demis Hassabis estimated five to ten years to AGI; he subsequently cited three to five; in a Wall Street Journal Magazine profile published this week, he gives two to three. Azhar's read is that this matters precisely because Hassabis is one of the field's more measured voices and is calibrating against what he sees inside his own lab. The caveat Azhar states and then sets aside is the right one: the definition of AGI is contested and arguably not useful, but the threshold of capability being marked out is real either way. Note the dating honestly — the WSJ Magazine interview ran earlier in the week, around 21–22 October, and what landed on Sunday was the framing rather than the quote.

So what: Treat the derivative, not the level. A forecaster cutting his own estimate twice in nine months is giving you information about the rate of surprise inside a frontier lab, which is more useful than the number he lands on — and it is the exponential-literacy point in a sentence: the people closest to the curve keep being surprised by it in the same direction. The planning implication is unglamorous: shorten the horizon on which you assume a capability gap protects a product. If your moat is a task frontier models cannot yet do, re-test that assumption quarterly rather than annually, and write the re-test into the roadmap now.

The internet under attack, AGI compression & the next oil shock that never will be... #607
Hassabis Cuts AGI Window to Two to Three Years While Safety Isn't Ready

A promised feature film arrived as seven minutes, which is the most useful vendor-claim datapoint of the weekend

TechCrunch's Anthony Ha wrote on Sunday afternoon about the first instalment of Grok's AI-generated Odyssey — seven minutes, billed as "Episode 1 of our Odyssey series," posted by the Grok Imagine account on X on Friday 23 October and made with Grok Imagine in collaboration with Wonder Studios and director Dev Walker. About three months ago Elon Musk said Grok would produce a full-length "historically accurate" film of the Odyssey. Ha's assessment is worth reporting in both directions: he found the footage better than he expected and better than other AI video Musk has posted, with some shots he might have taken for real actors on a real sea — and also airless, with stilted dialogue, an overlong voiceover backstory and no emotional stakes, something he would have skipped like a video-game cutscene.

So what: Keep this one as a calibration anchor rather than a story. The gap between "full-length film" promised in July and seven minutes delivered in October is a clean, dated, publicly checkable instance of the thing that makes vendor roadmaps unusable for planning — and the same frame applies to every agent demo you are shown this quarter. Ask what shipped, at what length, on what date, against what was promised. The constructive half is real too: shot quality crossed a line that a sceptical reviewer noticed, which means the binding constraint on AI video has moved from pixels to narrative, and narrative is a craft problem rather than a compute problem.

Grok releases the first seven minutes of its AI-generated 'Odyssey'
Elon Musk's Grok Releases First AI-Generated 'Odyssey' Episode

The open-versus-closed fight is the wrong axis for the cyber story, and the person who argued the open side is the one saying so

Interconnects published a Sunday podcast in which Nathan Lambert and Florian Brand worked through the month's security news, and the useful part is Lambert marking his own position to market. He has been arguing that releasing an open-weight Mythos-class model in April would have been fine; he now says what he failed to price in was the vulnerability trend line itself, which was already exponential before any of these releases. His conclusion: "none of this seems like an open-closed model story," it is ecosystem-dependent, and "the cost and ability to do cyberattacks is going way up, but the institutions that are doing them, that doesn't necessarily 10X their output." Brand adds the supply-side observation from the defensive side — that access to the labs' defensive programmes was the real bottleneck, which pushed researchers onto weaker open models, and that OpenAI's defensive programme is now straightforward to join with ID verification and hardware keys while the offensive one remains hard. Both note that the PHP vulnerability at the centre of the week was already patched when it was disclosed — Lambert says the fix was in progress when it leaked — and that mass exploitation has not materialised. They are also blunt about the commercial incentive working in defenders' favour: "there is so much money in fixing things."

So what: If you are setting AI security policy, stop spending the argument on model licences and spend it on two things you control: patch latency and credential hygiene. The week's two self-inflicted incidents make the point — LiteLLM traced its second compromise to an API key committed to GitHub five months earlier, and small firms are still running end-of-life PHP. Apply for the labs' defensive programmes now rather than later; the asymmetry Brand describes, easy defensive access and hard offensive access, is exactly the asymmetry you want to be on the right side of, and it is available by filling in a form.

Podcast: Kimi K3.1 Pro, the new cyber reality, and Muse musings

Every edition, with the full archive: https://excelsiorgroup.ai/insights/signal/