Where this is going

A page of instructions, because an appendix that is thrown away every quarter needs a different reading habit from the numbered parts. The slide titles are the durable layer: each is a way of thinking about the question, and it should survive several issues. The boxes are the perishable layer: every number carries the date it was checked and a source in the notes that was opened, not skimmed. Treat them as a photograph of October 2026. The word used throughout is expectation, not prediction: an expectation is where the evidence points, stated plainly enough that you can plan against it and discover cheaply when it is wrong. A prediction is a bet, and this series does not make them. The five content slides follow in order: what the people building the thing say, what the neutral trackers measure, what the money looks like, one expectation per numbered Part for the next twelve months, and the six questions that have not changed since Part I, with the maintenance schedule for the rest of the series.

Start with what the people building the thing say, because everyone in the room has read a headline about it and the headlines strip the hedges. Dario Amodei’s January 2026 essay says powerful AI “could be as little as 1–2 years away, although it could also be considerably further out”, and asks the reader to suppose a “country of geniuses in a datacenter” materialises somewhere “in ~2027”. On the Dwarkesh podcast in February he put the chance of that within ten years at 90 percent, said his hunch for the near case was “one to two, maybe more like one to three” years, and added that he “wouldn’t be surprised” to be off by a year or two. Sam Altman, in a TIME profile published 26 August 2026, said OpenAI would have an internal system it calls AGI by the end of the year, built on the Astra model family; the company’s charter defines AGI as “highly autonomous systems that outperform humans at most economically valuable work”, and there is no agreed external test for that. Demis Hassabis told the Wall Street Journal in October that his estimate “keeps moving closer to the present”, now two to three years; at Davos in January he had said five to ten, and that one or two more breakthroughs might be needed. Three things to notice. Each man’s company is valued on these claims. That does not make them wrong, but it is a reason to read the hedges, which are real and which headlines remove. The definitions differ: Hassabis means the full cognitive range of a human brain, OpenAI means economic output, Amodei describes a cluster of geniuses rather than one mind. And there are serious dissenters; Yann LeCun said at the same Davos that “we’re never going to get to human-level intelligence by training LLMs or by training on text only”. The durable rule is on the slide. Do not predict. Take the labs’ dates as a planning assumption, test your plans against the world where they are right and the world where they are five years early, and make sure neither bankrupts you.

Now the people with no product to sell. METR, a non-profit evaluator, measures the longest software task, in hours of human work, that the best model can finish half the time. Its March 2025 paper found that figure doubling about every seven months since 2019; its January 2026 re-fit, Time Horizon 1.1, found 131 days (confidence interval 107 to 161) for models since 2023 and 89 days for models since 2024. The highest reading it has published is Claude Opus 4.6 at around 14.5 hours in February 2026, with a 95 percent interval of 6 to 98 hours, which METR itself called “extremely noisy because our current task suite is nearly saturated”. By May it had added a warning that measurements above 16 hours are unreliable, and in September it marked the page as no longer actively updated. The caveats are METR’s own: error bars of roughly a factor of two, a suite that is almost entirely software and research engineering, horizons forty to a hundred times shorter on tasks that mean driving a screen, and an 80-percent-reliable horizon that trails far behind the headline. Epoch AI tracks the inputs. Its dashboard, refreshed 5 February 2026, has training compute for frontier language models growing five times a year since 2020 and the cost of a frontier run growing about three and a half times a year; the same group documented in Part IX that the price of buying a fixed level of capability falls by an order of magnitude or more each year. Both are true at once because the frontier keeps moving while last year’s frontier gets cheap. The third measurement is about measurement itself. ARC-AGI-3 launched in March 2026 with every frontier model under one percent; by September GPT-6 Astra scored 62.7 percent on the organisers’ standard harness, and 99.9 percent on OpenAI’s own adapter, which keeps reasoning state between steps. Same weights, different scaffolding, different score, and ARC Prize’s co-founder wrote that “we lack evidence to call this AGI yet”. Benchmarks now live months. A beaten benchmark proves the test is beatable, which is why the only evaluation that matters is yours.

The money question gets asked at every forum, so here it is handled like an adult: three numbers, both arguments, and what it means for someone buying rather than investing. The spending first. As guided after their most recent results, Amazon plans about 220 billion dollars of 2026 capital expenditure, Alphabet 195 to 205 billion, Microsoft about 175 billion for the calendar year (lower than an earlier 190 because of a lease-accounting change, not a cut), and Meta 135 to 145 billion, raised on 21 October. Call it 735 billion for four companies, against about 410 billion in 2025. Microsoft, Alphabet and Amazon report again in the last days of October, so this figure will move within days of this issue. The revenue next. OpenAI told investors in October that its annualised revenue was approaching 50 billion dollars, the Financial Times reported; a 70 billion figure that circulated in September turned out to be investors’ attempt at an apples-to-apples comparison with Anthropic, which counts sales through cloud partners and OpenAI does not. Anthropic’s S-1, public on 26 October, shows 16.3 billion of revenue in the first half of 2026 against 1.16 billion a year earlier, an operating profit of 974 million in the second quarter, and a headline net loss of 80.7 billion that is almost entirely a non-cash revaluation of convertible notes. The third number is the one buyers should enjoy: Ramp’s index of its business customers shows token volume up 52 percent in the four weeks to 11 October while spend rose 23 percent and the blended price fell to 64 cents per million tokens. The bubble case is Gartner’s arithmetic, published 14 October: data-centre owners need about 8.5 trillion dollars of cumulative revenue by 2030 to cover announced spending, financing and a 12 percent return, against at most 3.6 trillion of AI revenue it can see, leaving more than 3.5 trillion of spend exposed to impairment. Add the circularity of clouds investing in labs that commit hundreds of billions back to clouds, and a market that fell on a single revenue footnote on 8 October. The case against is that demand is measured rather than promised, that the four big spenders fund it from cash flow and are, in Gartner’s own view, the least exposed, and that the gap can close from either side. For a buyer the answer does not depend on who is right: negotiate short, pin versions, keep a fallback that passes your eval, and keep the layer from Part IX.

One line per Part, each the thing most likely to have changed by next autumn, and each labelled as an expectation rather than a prediction, because the difference is the whole point of this deck. Part II: open models of about thirty billion parameters, Meta’s Apache-licensed Muse Glimmer among them, already run on a good laptop and match where the frontier stood a year earlier; expect more of your jobs to pass the local test, and expect licences to keep changing. Part III: agent features will arrive inside software you already pay for, mostly unasked, and Gartner’s June 2025 forecast that more than 40 percent of agentic projects will be cancelled by the end of 2027 still stands; run the sniff test on each. Part IV: longer context, more hidden reasoning billed as output, and labs using models to build the next model; the mechanism does not change, the bill does. Part V: the chat box becomes an operator that drives a browser and desktop, and free consumer chat carries advertising, which ChatGPT has done for free and Go users since February 2026. Part VI: standard plugs, with MCP under the Linux Foundation’s Agentic AI Foundation since December 2025, and the ARC result on slide three as the proof that the harness, not just the model, sets the score. Part VII: McKinsey’s August 2026 survey found about one in five respondents saying AI operating costs, token costs included, constrain their use, which makes evals a budget tool. Part VIII: after the EU’s Digital Omnibus, two new prohibitions and labelling duties for generative systems already on the market apply from 2 December 2026, and Annex III high-risk obligations moved from August 2026 to 2 December 2027. Part IX: Google’s own price page doubles Gemini 4 Argon to four and twenty dollars per million tokens on 1 December 2026, open models keep undercutting, Anthropic’s S-1 is public, and OpenAI confirmed a confidential filing in June while Altman told Fortune in September it will not list in 2026. Part X: Agility’s Digit goes to two more GXO warehouses from early 2027 after 100,000 totes at the first, Boston Dynamics says Atlas reaches customers beyond Hyundai and DeepMind in early 2027, and Tesla began what Musk called “early production” of Optimus in Fremont on 21 October; still fenced floors, still no published reliability figures. Parts XI and XII, from the same McKinsey survey: about two in ten organisations scaling coding agents, 39 percent expecting AI to reduce headcount against 14 percent that saw it happen, and 37 percent reporting any EBIT impact, unchanged for a year.

The last slide is the one that does not get re-issued. Six questions have run through every part of the series, and they are the things to hold onto when the names on the other four slides have gone stale. What is the job: Parts I and XII both insist that the question is never “where can we use AI” but which inbox, which queue, which decision, written down well enough that a stranger could do it. Who holds the weights: Parts II and IX, a file you keep or a key someone else can revoke, with the licence read, the version pinned and a fallback that passes your eval. What is the agent allowed to touch: Parts III, VI and VIII, where prompts are suggestions and permissions are walls. Who is awake when it fails: a named human, a log and a kill switch, because “the vendor” has never once been the answer. What does it cost per job: Parts VII and IX, the only number that survives a price change. And who owns the output: you do, as Air Canada discovered in a British Columbia tribunal in February 2024, for every word, action and picture made under your name. The right-hand panel is the maintenance schedule. This deck is re-issued every quarter, and with it the slides in the main series that carry a check date: Part IX’s price board and map, Part X’s bodies board, Part III’s platform rules and build-or-buy slide, Part VIII’s regulation dates, Part II’s hardware and model names, Part I’s map of who is on the field. Thirty minutes a quarter: re-run the three-model eval, re-read the board, re-read the licence, update one slide. Then the closing line, which is the series’ own: the names will be wrong by February; the questions will not.