Open predictions
78 predictions by Gary Marcus and Ed Zitron are unresolved and still inside the window they gave, as of 2026-10-29. When a window closes, the claim is checked again. A closed window does not make a prediction wrong.
Windows that end in 2026 (12)
“Human domestic robots like Optimus and Figure will be all demo and very little product.”
Our reading: In 2026, humanoid domestic robots such as Optimus and Figure will be mostly demos with very little actual product.
Counts as supported if: By 2026-12-31, consumer home deliveries of Optimus, Figure and similar humanoids are absent or tiny (pilot or early-access scale), while demos are plentiful.
Counts as contradicted if: During 2026, reliable sources document a substantial product: thousands of humanoid robots sold and delivered to consumer homes and used there regularly.
“Work on new approaches such as world models and neurosymbolic will escalate.”
Our reading: In 2026, work on new approaches such as world models and neurosymbolic AI will escalate.
Counts as supported if: Comparable evidence shows a broad increase in research and development activity during 2026 over 2025, including the named approaches.
Counts as contradicted if: Comparable evidence covering 2026 shows that activity across the named approaches remained broadly flat or declined.
“We won’t get to AGI in 2026 (or 7).”
Our reading: AGI will not be achieved in 2026 or 2027.
Counts as supported if: By 2027-12-31, no AI system has been broadly accepted by independent experts as AGI.
Counts as contradicted if: Before 2028, a system is broadly accepted as AGI, for example confirmed by an independent verification process and widely acknowledged by independent researchers.
“there will be tons of lab demos but getting these robots to work in people’s homes will be very very hard, as Rodney Brooks has said many times.”
Our reading: There will be many lab demos of humanoid robots in 2026, but getting them to work in homes will be very hard.
Counts as supported if: By the end of 2026, demos of humanoid home robots are plentiful, while real home use is rare, limited, heavily teleoperated or shown to be unreliable in independent reviews.
Counts as contradicted if: During 2026, humanoid robots are shown in independent reviews to work reliably and autonomously in many consumer homes, or demos are few.
“No country will take a decisive lead in the GenAI “race”.”
Our reading: No country will take a decisive lead in the generative AI race in 2026.
Counts as supported if: Through the end of 2026, independent comparisons (for example the Stanford AI Index or leaderboards) show the best models from leading countries within a small margin of each other, or trading the lead.
Counts as contradicted if: By the end of 2026, independent comparisons show one country's models clearly and durably ahead of all others by a wide margin, widely described as a decisive lead.
“Or even how he eventually walked that back to 2026 (which also won’t happen).”
Our reading: Elon Musk's promise of a million Tesla robotaxis, pushed back to 2026, will not be fulfilled in 2026.
Counts as supported if: Reliable records covering the completed deadline establish that Tesla did not reach one million operating robotaxis during 2026.
Counts as contradicted if: Reliable deployment records establish that Tesla reached at least one million operating robotaxis during 2026.
“Developer Mortensen claims the entire project will be completed by October 2026, which it obviously, blatantly won’t.”
Our reading: Stargate Abilene will not be fully completed by October 2026.
Counts as supported if: Dated records establish that at least one of the eight buildings remained unfinished or unenergized at the end of October 31, 2026.
Counts as contradicted if: Dated records establish that all eight buildings were completed and energized on or before October 31, 2026.
“I’m making the call that by the end of 2026, a majority of AI services will move some or all of their customers to token-based billing as they reckon with the true costs of running AI models.”
Only if: some or all of their customers
Our reading: By the end of 2026, most AI services will move some or all customers to token-based billing.
Counts as supported if: Dated billing records establish that more than half of a defensibly defined service population transitioned at least some existing customers to token-based billing between the post and December 31, 2026.
Counts as contradicted if: At the deadline, a sufficiently complete population count establishes that half or fewer made a qualifying transition.
“OpenAI Represents 70%+ Of Microsoft’s AI Revenue and 80%+ Of Its AI GPU Compute Capacity, Creating The Illusion Of Growth That’s Dependent On A Company That Will Lose $25 Billion+ In 2026”
Counts as supported if: Matched-period records establish both stated Microsoft concentration thresholds, and OpenAI's realized calendar-2026 net loss reaches at least $25 billion. Historical and contemporaneous readings must be distinguished.
Counts as contradicted if: A reliable matched-period record disproves either concentration threshold, or OpenAI's finalized calendar-2026 net loss is below $25 billion. A cash-burn projection cannot establish or contradict net loss.
“Prediction: Nobody knows when this will all collapse, but 2026 will be remembered in hindsight as the year in which retail investors and index funds were left holding the bag.”
Only if: collapse timing unknown
Counts as supported if: Ownership and investment-flow records establish substantial 2026 exposure for both retail investors and index funds, with material subsequent losses attributable to that exposure and evidence supporting the claimed transfer of downside to those holders.
Counts as contradicted if: Complete records for the 2026 episode and its subsequent resolution show no substantial transfer of such exposure to both groups, or establish that the relevant holdings did not leave them bearing material losses. A short-term rally is insufficient.
“Buying GPUs and building data centers sort of feels like Amazon Web Services, even though the $765 billion that big tech will spend in 2026 will be more than ten times Amazon’s combined capex during the period where AWS was being built.”
Our reading: Big tech will spend $765 billion in 2026, more than ten times Amazon's combined capex during the AWS build-out.
Counts as supported if: Actual 2026 spending for a consistently defined cohort and expenditure measure is approximately $765 billion and exceeds ten times comparable Amazon capital expenditure over 2003–2017.
Counts as contradicted if: Comparable completed-year spending materially differs from $765 billion, or the ratio to the verified Amazon baseline is ten or less.
“Hyperscalers will have sunk over $1.3 trillion dollars into generative AI by the end of 2026”
Counts as supported if: From March 2022 to 2026-12-31, the four companies sunk over $1.3 trillion into generative AI, on the measure the speaker meant.
Counts as contradicted if: From March 2022 to 2026-12-31, the four companies sunk $1.3 trillion or less into generative AI, on the measure the speaker meant.
Windows that end in 2027 (6)
“If there exist AI systems that can perform 8 of the 10 tasks below by the end of 2027, as determined by our panel of judges, Gary will donate $2,000 to a charity of Miles’ choice; if AI can do fewer than 8, Miles will donate $20,000 to a charity of Gary’s choice.”
Only if: as determined by a panel of judges with broad discretion; success judged on technical feasibility at modest per-task cost (say $1M compute + 10 person-weeks), even if no system has actually attempted a task; tasks not performed by 10 completely separate systems, though some task tailoring is allowed; excludes tasks involving physical action in the world; the post says the 10:1 odds reflect Miles' confidence and readers should not infer Gary's beliefs from the odds
Counts as supported if: The mutually agreed panel determines that fewer than eight of the ten specified tasks were technically feasible by December 31, 2027, under the bet’s cost, tailoring and system constraints.
Counts as contradicted if: The mutually agreed panel determines that at least eight of the ten specified tasks were technically feasible by December 31, 2027, under those constraints.
“I predict that such counterexamples will still be easy to find at the end of 2027 because of limitations inherent in LLMs.”
Only if: because of limitations inherent in LLMs
Counts as supported if: Testing at the end of 2027 readily finds clear compositional failures in contemporary leading LLM-based generators, and comparative technical evidence supports the attribution to inherent LLM limitations.
Counts as contradicted if: Broad, documented end-of-2027 testing shows that such failures are no longer readily elicited, or strong technical evidence refutes the asserted inherent cause of the relevant failures.
“AGI ain’t actually coming in 2027, and even Kokotajlo, known for his aggressive timelines, is now with me in thinking we probably won’t see AGI this decade.”
Our reading: AGI will not arrive in 2027.
Counts as supported if: After December 31, 2027, a sufficiently comprehensive public record shows that no system meeting the stated general-capability criterion arrived by the deadline.
Counts as contradicted if: A dated release or demonstration establishes AGI by December 31, 2027, with independent evidence of broad capabilities and recognition beyond the developer's assertion.
“By the beginning of 2027, Stargate Abilene will be obsolete, as will any and all data centers filled with Blackwell GPUs, as will any and all data centers being built today.”
Counts as supported if: Evidence at the deadline establishes commercial obsolescence of Abilene and the entire stated facility cohort, including an inability to serve economically useful AI workloads.
Counts as contradicted if: At least one facility within the stated cohort remains commercially useful for AI workloads at the deadline, supported by actual deployment, customer use and operating evidence.
“By the end of 2027, big tech will have sunk $2 trillion into AI capex, with very little to show for it.”
Our reading: By the end of 2027 big tech will have spent $2 trillion on AI capex with very little to show for it.
Counts as supported if: By December 31, 2027, comparable records establish at least $2 trillion of cumulative AI capital expenditure by the four named companies, together with clearly small attributable financial returns relative to that investment.
Counts as contradicted if: Comparable records establish cumulative AI capital expenditure below the stated milestone at the deadline, or substantial attributable financial returns that clearly defeat the ordinary reading of 'very little'. Borderline returns cannot be classified using an invented cutoff.
“As mentioned, this is going to begin in earnest in 2027, and expand dramatically every year following (though I doubt we will make it that far).”
Only if: though I doubt we will make it that far
Counts as supported if: Comparable payment records show a substantial ramp in 2027 relative to 2026 and substantial annual increases thereafter. A finite record can support only the years observed, not an indefinite continuation.
Counts as contradicted if: Comparable records show no substantial 2027 ramp, or a subsequent completed year clearly fails to show the asserted substantial increase.
Windows that end in 2029 (7)
“In 2029, AI will not be able to reliably construct bug-free code of more than 10,000 lines from natural language specification or by interactions with a non-expert user.”
Only if: from natural language spec or non-expert interaction; gluing existing library code doesn't count
Counts as supported if: Evaluations during 2029, covering leading available systems and varied qualifying tasks, consistently find that systems cannot reliably produce correct new programs exceeding 10,000 substantive source lines under either permitted input route. Mere absence of a public demonstration is insufficient.
Counts as contradicted if: A system repeatedly produces correct new programs exceeding 10,000 substantive source lines from natural-language specifications or non-expert interaction, with independently assessed correctness and without counting library assembly as the implementation.
“In 2029, AI will not be able to read a novel and reliably answer questions about plot, character, conflicts, motivations, etc.”
Only if: key will be going beyond the literal text
Counts as supported if: Evidence covering the end of 2029 shows that leading available systems still fail to answer questions dependably about whole novels, including questions requiring inference about plot, characters, conflicts and motivations.
Counts as contradicted if: At least one AI system reliably answers questions across varied whole novels, including nonliteral questions, under a credible evaluation with human-validated answers.
“In 2029, AI will not be able to take arbitrary proofs from the mathematical literature written in natural language and convert them into a symbolic form suitable for symbolic verification.”
Only if: from the mathematical literature
Counts as supported if: Broad evaluations during 2029 find that leading systems still cannot faithfully formalize independently selected natural-language proofs from diverse mathematical literature, with recurring substantive failures rather than merely occasional errors. Absence of a demonstration alone is insufficient.
Counts as contradicted if: Reproducible evaluations establish broad successful conversion of independently selected literature proofs into checked symbolic proofs, preserving the original mathematical content and reasoning without substantial expert completion.
“In 2029, AI will not be able to work as a competent cook in an arbitrary kitchen (extending Steve Wozniak’s cup of coffee benchmark).”
Our reading: In 2029, AI will not be able to work as a competent cook in an arbitrary kitchen.
Counts as supported if: Evidence covering the end of 2029 shows that available AI robots still cannot consistently complete ordinary cooking tasks in varied unfamiliar kitchens without substantial human assistance or robot-specific installation.
Counts as contradicted if: An AI robot repeatedly prepares ordinary meals competently and safely across varied unfamiliar kitchens without teleoperation, task completion by people, or bespoke kitchen installation.
“In 2029, AI will not be able to watch a movie and tell you accurately what is going on (what I called the comprehension challenge in The New Yorker, in 2014).”
Our reading: In 2029, AI will not be able to watch a movie and accurately describe what is going on.
Counts as supported if: Evidence covering the end of 2029 shows that leading available systems still cannot accurately explain full movies across a representative evaluation of events, characters, conflicts and motivations.
Counts as contradicted if: At least one AI system demonstrates accurate full-movie comprehension across varied movies under a credible evaluation that tests the stated narrative elements.
“There is simply no space in these projections to absorb that $300 billion, and from what I can tell, by 2029, OpenAI will have actually burned more than $290 billion, assuming that it survives that long, which I do not believe it will.”
Our reading: OpenAI will not survive until 2029.
Counts as supported if: OpenAI permanently ceases business operations by the deadline, with no continuing business successor.
Counts as contradicted if: OpenAI continues operating through the deadline.
“While Anthropic claims it will be profitable by then, I do not believe it will be, nor is it profitable at this point outside of financial engineering.”
Our reading: Anthropic will not be profitable by 2029, despite its claims.
Counts as supported if: Comparable annual results through 2029 show that Anthropic has not reached positive company-wide profit after expenses.
Counts as contradicted if: Comparable results establish positive company-wide profit for a completed full year by the end of 2029.
Windows that end in 2030 (2)
“Let’s get specific: 2023, Microsoft, Google, Amazon, and Meta have spent a little over $850 billion in capex, mostly hoarding NVIDIA GPUs that will be near-to-completely obsolete by 2030.”
Our reading: The NVIDIA GPUs big tech bought with ~$850B capex since 2023 will be nearly completely obsolete by 2030.
Counts as supported if: By the deadline, fleet and market evidence shows that almost all NVIDIA GPUs in the specified purchase cohort have lost productive commercial usefulness, with negligible economically useful deployment or resale demand.
Counts as contradicted if: At the deadline, a substantial portion of that cohort remains productively deployed or retains meaningful commercial demand and value.
“This may seem like the cheat code to get around all these horrifying rates, except CoreWeave, per analysts at UBS, needs to raise $102 billion in debt through 2030, which means it will have to keep raising on its own two feet.”
Only if: $102 billion need is a UBS analysts' estimate
Counts as supported if: Evidence through the deadline establishes the claimed debt-financing requirement and continued borrowing materially dependent on CoreWeave's own credit, rather than exclusively on stronger counterparties.
Counts as contradicted if: Evidence establishes that the claimed financing program can be completed without continued borrowing on CoreWeave's own credit, or establishes a materially different required debt amount on the same forecast basis.
Windows that end in 2033 (1)
“I hereby bet publicly that AI will not fully displace radiologists (as opposed to merely augmenting them) before the year 2033.”
Our reading: AI will not fully displace radiologists, as opposed to augmenting them, before 2033.
Counts as supported if: At the end of 2032, credible workforce and clinical-practice records establish that radiologists remain a practicing profession and have not been fully displaced by AI.
Counts as contradicted if: Before January 1, 2033, independently verified deployment and workforce records establish that AI has fully replaced radiologists as a profession.
Windows that end in 2043 (2)
“In retrospectives written in 2043, intellectual historians will conclude that there was an initial overemphasis on large language models, and a gradual but critical shift of the pendulum back to more structured systems with deeper comprehension.”
Counts as supported if: A representative body of intellectual historians' AI retrospectives written in 2043 endorses both specified judgments.
Counts as contradicted if: A representative body of such retrospectives explicitly rejects one or both judgments. Mere silence or failure to find retrospectives does not establish contradiction.
“In retrospectives written in 2043, intellectual historians will conclude that there was an initial overemphasis on large language models, and a gradual but critical shift of the pendulum back to more structured systems with deeper comprehension.”
Only if: intellectual historians
Counts as supported if: Relevant retrospectives written in 2043 characteristically conclude both that LLMs initially received excessive emphasis and that a gradual, consequential shift toward more structured systems with deeper comprehension followed.
Counts as contradicted if: A sufficiently broad body of relevant retrospectives written in 2043 characteristically rejects that account or gives a materially incompatible account of the field's development.
Windows counted from the post date (22)
“So I don’t think we will waste more than 5 years before a substantial fraction of AI effort is looking outside the limited streetlight that currently are sucking up so much mental energy.”
Our reading: Within 5 years, a substantial fraction of AI effort will look beyond the current LLM-focused approaches.
Counts as supported if: By January 22, 2028, representative research-output, staffing, or funding records establish an unambiguously substantial share of AI effort pursuing approaches beyond the contemporary LLM-centered focus.
Counts as contradicted if: Representative records through the deadline show that such work remains marginal under reasonable measures and classifications of AI effort.
“Driver-assist (a la Tesla) with humans in the loop may well stay, occasionally causing fatal accidents, given its lack of reliability, but the idea of using your car moonlighting as a self-driving taxi is not happening this decade, maybe not even next.”
Only if: maybe not even next decade
Our reading: Personal cars moonlighting as self-driving taxis will not happen this decade, and maybe not next decade either.
Counts as supported if: After December 31, 2029, sufficiently comprehensive operational evidence establishes that no qualifying personal-car taxi service operated during the forecast interval.
Counts as contradicted if: Dated records establish that privately owned personal cars actually earned passenger fares through a driverless taxi service before January 1, 2030.
“My offer to bet him a million dollars on this nonsense stands.”
Counts as supported if: At 2026-10-30, AI systems still cannot do many tasks humans can do, as shown by public capability evaluations.
Counts as contradicted if: By 2026-10-30, credible evidence shows an AI system that can do anything any human can do.
“That is not going to magically change next year, at a time when AI still can’t reliably solving river-crossing problems.”
Our reading: In 2026, AI will still not be choosing its own research goals in place of human scientists.
Counts as supported if: Evidence covering 2026 shows that leading AI-for-science systems continue to depend on human scientists to choose substantive research objectives, despite automating hypotheses, experiments, or other research steps.
Counts as contradicted if: Documented research records show AI repeatedly choosing and pursuing substantive scientific objectives in place of human scientists, with enough detail to distinguish independent research direction from work within a human-assigned objective.
“As a reminder, OpenAI has leaked that it’ll burn $115 billion in the next four years, and based on my estimates, it needs to raise more than $290 billion in the next four years based on its $300 billion deal with Oracle alone.”
Only if: based on his estimates; based on the Oracle deal alone
Counts as supported if: A documented funding reconciliation establishes that meeting the Oracle obligations requires more than $290 billion of additional external financing within the four-year interval.
Counts as contradicted if: A complete reconciliation establishes that the Oracle obligations require no more than $290 billion of additional external financing within that interval.
“AGI is NOT in fact coming in Trump’s second term, and never was.”
Our reading: AGI will not arrive during Trump's second term (ending January 2029).
Counts as supported if: After the term ends, sufficiently comprehensive evidence establishes that no qualifying AGI system arrived during it.
Counts as contradicted if: A dated release or demonstration establishes a qualifying AGI arrival during the second term, before noon Eastern time on January 20, 2029.
“Both companies burn billions of dollars, and require an infinite amount of venture capital to keep alive at a time when the Saudi Public Investment Fund is struggling and the US venture capital system is set to run out of cash in the next year and a half.”
Counts as supported if: Reliable system-wide records establish exhaustion of deployable US venture-fund capital within the forecast interval.
Counts as contradicted if: At the deadline, material deployable capital remains and records show no earlier system-wide exhaustion during the interval.
“I don’t expect my predictions to be as on target this year as last, for a happy reason: across the field, the intellectual situation has gone from one that was stagnant (all LLMs all the time) and unrealistic (“AGI is nigh”) to one that is more fluid, more realistic, and more open-minded.”
Counts as supported if: Independent scoring of fixed, comparable prediction sets establishes a lower 2026 accuracy rate than the 2025 rate. Supporting the stated explanation additionally requires evidence connecting the field's changes to the forecasting errors.
Counts as contradicted if: Independent scoring of fixed, comparable prediction sets establishes a 2026 accuracy rate equal to or higher than the 2025 rate.
“AI can help in the search for candidate drugs, but we still need to test them. There is zero chance that we double human life span in the next decade.”
Our reading: There is zero chance that human life span is doubled in the next decade.
Counts as supported if: After March 8, 2036, suitable human evidence shows that the specified lifespan measure did not double during the forecast window.
Counts as contradicted if: Reliable human evidence establishes a doubling of typical lifespan relative to the same March 2026 baseline by March 8, 2036.
“A rack of Vera Rubin or Blackwell GPUs will cost as much to run in five years as they do today, as will an incomplete data center cost just as much to finish construction, connect to the grid or acquire behind-the-meter (IE: generators) power for.”
Counts as supported if: Comparable June 2031 records show no material reduction from June 2026 in both rack running costs and the costs of completing and powering comparable unfinished facilities.
Counts as contradicted if: Comparable June 2031 records show materially lower running costs for the specified racks or materially lower completion and power-provision costs for comparable unfinished facilities. Divergent results across components would require a mixed assessment.
“We won’t see AGI this decade, and we not even close, exactly for the reasons that he laid out.”
Only if: AGI as a system with all human cognitive capabilities, including creativity and physical intelligence
Counts as supported if: By the end of 2029, credible assessments of available systems establish that none meets the endorsed AGI definition.
Counts as contradicted if: A system achieved by December 31, 2029 demonstrably meets the endorsed definition, with independent verification across cognitive, creative, and physical capabilities.
“We won’t see AGI this decade, and we not even close, exactly for the reasons that he laid out.”
Only if: exactly for the reasons Hassabis laid out (no genuine invention across domains, no physical intelligence)
Counts as supported if: At the end of December 31, 2029, sufficiently comprehensive public records establish that no system meeting the endorsed AGI definition was publicly demonstrated during the decade.
Counts as contradicted if: A credible, dated public demonstration establishes that a system met the entire endorsed AGI definition by December 31, 2029.
“These funds will be insufficient to cover Anthropic’s costs, as will Anthropic’s cash flow, meaning that it will have to raise at least another $200 billion in the next year.”
Counts as supported if: Between 2026-06-08 and 2027-06-08, Anthropic raised at least $200 billion of new external funding, and its disclosures show it needed the money to cover its costs.
Counts as contradicted if: By 2027-06-08, Anthropic had raised less than $200 billion of new external funding and its disclosures show it could cover its costs.
“These funds will be insufficient to cover Anthropic’s costs, as will Anthropic’s cash flow, meaning that it will have to raise at least another $200 billion in the next year.”
Counts as supported if: Anthropic obtains at least USD 200 billion of additional financing by June 8, 2027, and cash-flow records establish that the baseline funds and operating cash flows could not cover its costs without that amount.
Counts as contradicted if: At the deadline, Anthropic has met the relevant costs with less than USD 200 billion of additional financing, or an irreversible earlier settlement establishes that less financing suffices for the entire period.
“The $122 billion funding round from March will be insufficient to cover these costs, and it will require, at the very least, another $250 billion in funding by the end of the year.”
Counts as supported if: Financial records establish that the March round cannot cover the referenced costs and that OpenAI requires at least USD 250 billion of additional external financing by December 31, 2026, after accounting for existing liquidity, operating cash receipts, and payment timing.
Counts as contradicted if: Complete records establish that OpenAI can meet its obligations through December 31, 2026 with less than USD 250 billion of additional financing beyond the March round. Merely raising less, without establishing ability to pay, is insufficient.
“Based on my own reporting on its audited financials from 2024 and 2025, OpenAI will need to raise funding at least three more times in the next decade.”
Our reading: OpenAI will need to raise funding at least three more times in the next decade.
Counts as supported if: At least three distinct additional financings occur within the horizon, with financial records establishing that they were necessary.
Counts as contradicted if: Complete records through the deadline establish fewer than three necessary additional financings.
“Absolutely absurd! I hereby offer @elonmusk a million dollar bet against his prediction that Optimus will be better than the best humans in surgery by the end of the decade.”
Our reading: Optimus will not be better than the best human surgeons by the end of the decade (Marcus bets $1M against it).
Counts as supported if: At the deadline, sufficiently comprehensive evidence shows that Tesla Optimus has not achieved the specified surgical superiority.
Counts as contradicted if: By December 31, 2029, credible comparative clinical evidence establishes that Tesla Optimus performs surgery autonomously better than leading human surgeons across a representative range of procedures.
“There are only two companies that actually spend more than a few hundred million a year on AI compute — OpenAI and Anthropic — and both of them will lose tens of billions of dollars this year and more next year.”
Our reading: OpenAI and Anthropic will each lose tens of billions of dollars in 2026 and more in 2027.
Counts as supported if: Each company reports a comparable annual net loss of at least $20 billion for 2026 and a larger annual net loss for 2027.
Counts as contradicted if: Final comparable results show that at least one company fails either its 2026 minimum or its 2027 increase.
“AI might someday contribute as Hassabis says, but certainly not in the next decade.”
Our reading: AI will not make the disease-curing contributions Hassabis described within the next decade, though it might someday.
Counts as supported if: At the deadline, reliable clinical evidence establishes that AI has not delivered the broad disease-curing achievement described in the referenced interview.
Counts as contradicted if: By the deadline, clinical evidence establishes that AI materially contributed to demonstrated cures across the full disease scope described by Hassabis.
“Google, Amazon and Microsoft are currently spending hundreds of billions of dollars on capex to, for the most part, pull in around $440 billion in revenue in the next three years from Anthropic and OpenAI, as they represent more than 70% of their AI revenues in the next three years, and more than 48% of Google Cloud’s 2027 revenue.”
Only if: for the most part
Counts as supported if: Comparable customer-level accounts establish approximately $440 billion of realized revenue during the three-year interval, a share above 70% of aggregate AI revenue during that interval, and a share above 48% of Google Cloud revenue in calendar 2027. Evidence must also substantiate the current capital-spending scale and its claimed predominant purpose.
Counts as contradicted if: Comparable completed-period accounts establish failure of a numerical component, allowing for the unspecified approximation around $440 billion; or reliable evidence directly disproves the claimed predominant purpose of capital spending. A shortfall before the deadline does not qualify.
“Next year, at least $200 billion in compute costs are coming due.”
Our reading: At least $200 billion of OpenAI and Anthropic compute costs will come due in 2027.
Counts as supported if: Reliable records establish at least USD 200 billion of combined OpenAI and Anthropic compute bills falling due during calendar 2027.
Counts as contradicted if: Complete records establish that their combined compute bills falling due during calendar 2027 total less than USD 200 billion.
“Yet things get a little messier when you factor in non-cancelable contracts, which total $413 billion in Anthropic’s case across the next seven to ten years, guaranteeing financial strain at a time when future cashflows are far from guaranteed.”
Counts as supported if: Evidence establishes the stated commitments and shows material financial strain caused by meeting them during the forecast horizon.
Counts as contradicted if: Anthropic meets or extinguishes the relevant commitments without material financial strain through their final maturity.
Conditions we are watching (17)
Each of these comes due only if its condition happens first. A date in the condition is not the deadline for the prediction.
“And because he has become so tight with Silicon Valley he will also be closely tied to any AI-tinged economic debacle that happens on his watch.”
Only if: if an AI-tinged economic debacle happens on his watch
Counts as supported if: An AI-linked economic debacle occurs before 2029-01-20, and mainstream reporting and polling widely tie it to Trump and his closeness to the tech industry.
Counts as contradicted if: An AI-linked economic debacle occurs before 2029-01-20, but reporting and polling largely do not tie it to Trump.
“At its current rate of construction, Stargate Abilene will be fully built sometime in late 2027.”
Only if: at its current rate of construction
Our reading: At its current construction rate, Stargate Abilene will be fully built in late 2027.
Counts as supported if: Evidence establishes that the March 2026 construction rate continued and that the original eight-building campus first became fully built during October–December 2027.
Counts as contradicted if: With the unchanged-rate condition established, the campus first becomes fully built before October 2027 or remains unfinished after December 31, 2027.
“Simple questions like “where are the GPUs going?” and “how many actual GPUs have been installed?” are left unanswered as article after article gets written about massive, multi-billion dollar compute deals for data centers that won’t be built before, at this rate, 2030.”
Only if: at the current rate
Our reading: At the current rate, data centers in the announced multi-billion dollar compute deals will not be built before 2030.
Counts as supported if: For an identified project set whose referenced construction pace persisted, the relevant facilities or phases remain incomplete through December 31, 2029.
Counts as contradicted if: An identified in-scope facility or contracted phase is completed before January 1, 2030 while the stated construction-rate condition holds.
“Doubling that won’t be enough. Tripling it won’t be enough. In fact, to pay this off, these companies will need to be making over $100 billion each in AI revenue in the next year, because otherwise there is no covering these losses.”
Only if: to pay off over $900 billion in investments; assuming the best-case $70 billion combined current revenue
Counts as supported if: With the stated premises established, comparable investment-recovery accounts show that each company required more than $100 billion in annual AI revenue within the horizon, and that doubling or tripling the stated combined revenue was insufficient.
Counts as contradicted if: With the premises established, at least one company demonstrably recovers its relevant investment without exceeding $100 billion in annual AI revenue within the horizon, or matching accounts show recovery possible at the stated doubled or tripled combined revenue.
“The report also claims that OpenAI is “on track” to hit its goal of generating $30bn in revenue for 2026, although if it maintains these disastrous margins, it would end up losing $36.6bn.”
Only if: if OpenAI maintains -122% margins; if it hits $30bn revenue
Our reading: If OpenAI hits $30bn revenue in 2026 at its Q1 margins, it would lose about $36.6bn in 2026.
Counts as supported if: Full-year 2026 revenue about $30bn, adjusted margin about -122%, and an adjusted operating loss near $36.6bn.
Counts as contradicted if: Revenue about $30bn and margin about -122% held, yet the adjusted loss was far from $36.6bn.
“When I say that, I mean it absolutely has to do that otherwise none of the data center capex makes sense, and neither Anthropic nor OpenAI can pay their commitments.”
Only if: if AI does not generate over $2 trillion in annual revenue by 2030
Counts as supported if: Evidence establishes the revenue antecedent and shows that the referenced data-center investment is financially unjustified and that both Anthropic and OpenAI cannot meet the relevant payments when due.
Counts as contradicted if: With the revenue antecedent established, evidence shows that a covered data-center investment is financially viable or that either Anthropic or OpenAI can meet its relevant commitments. Assess the two consequences separately if evidence differs.
“Both Anthropic and OpenAI need to be making $10 billion or more in monthly revenue by Q1 2028, or their growth rates aren’t going to support their compute commitments.”
Only if: to support their compute commitments
Counts as supported if: After a company misses the revenue threshold, dated financial and contractual evidence shows that insufficient growth leaves it unable to fund its compute obligations.
Counts as contradicted if: A company misses the threshold but demonstrably retains and funds the relevant compute obligations through a credible payment and financing plan.
“To pay for this compute, Anthropic must meet its projected revenue of $174 billion a year by 2029.”
Only if: to pay for its compute commitments
Our reading: Anthropic must reach its projected $174 billion in annual revenue by 2029 to pay for its compute commitments.
Counts as supported if: Payment records and a complete cash-funding reconciliation establish that paying the identified compute obligations required at least USD 174 billion in annual revenue by the deadline.
Counts as contradicted if: Anthropic pays the identified obligations while remaining below USD 174 billion in annual revenue through the deadline, using other available resources, or binding payment and financing records establish that the threshold was unnecessary.
“Anthropic needs to be making over $100 billion in two years or it can’t afford its commitments”
Only if: to afford its commitments
Our reading: Unless Anthropic is making over $100 billion within two years, it cannot afford its commitments.
Counts as supported if: Evidence for the deadline establishes that annual revenue of $100 billion or less cannot fund the relevant payments even after available cash and financing are considered, and demonstrates why revenue exceeding that threshold is necessary.
Counts as contradicted if: Anthropic has annual revenue of $100 billion or less at the deadline while demonstrably retaining sufficient resources to meet the relevant commitments when due.
“AI cannot, under any circumstances, slow down. In a year, Anthropic and OpenAI’s businesses have to be roughly twice the size they are today, and then double again basically every year until 2029 or 2030.”
Only if: to sustain the AI industry's commitments
Counts as supported if: A material shortfall from approximately annual revenue doubling occurs and financial evidence shows that the resulting growth shortfall prevents the relevant commitments from being funded.
Counts as contradicted if: Either company materially misses the approximate doubling path yet demonstrably sustains its relevant commitments through adequate operating resources and financing.
“There is no avoiding the carnage to come, outside of there being somewhere in the order of ten to a hundred times the demand for AI compute by 2030 that exists today, which would require AI compute to be larger than the $779 billion that the software industry earns annually.”
Only if: unless AI compute demand is somewhere in the order of 10 to 100 times today's by 2030
Counts as supported if: Comparable measurements establish that demand failed to reach the required scale by the end of 2030, and the relevant AI investment cycle produces widespread severe failures and losses. The dollar implication additionally requires a documented conversion between demand and annual compute spending.
Counts as contradicted if: A completed relevant investment cycle avoids severe AI-sector collapse despite demand remaining below the necessary scale-up. The monetary implication could separately be contradicted by comparable demand and price measurements showing the claimed growth without exceeding the stated annual dollar benchmark.
“I doubt there will be a single magic moment that our descendants think of as the “Singularity”. If there is, it’s not one that’s going to happen in this decade.”
Only if: if there is a single magic moment thought of as the Singularity; I doubt there will be such a single moment
Counts as supported if: The antecedent is established, and credible historical recognition places the single qualifying event after December 31, 2029.
Counts as contradicted if: The antecedent is established, and credible evidence dates the qualifying event within July 28, 2026 through December 31, 2029.
“No matter how big, handsome, and amazing you think either of these companies are, their expected compute spend will require them to make as much revenue as Microsoft, Google and Amazon in the next four years, and if they don’t, hyperscalers will not meet analyst and investor expectations.”
Only if: if OpenAI and Anthropic don't reach hyperscaler-level revenue
Counts as supported if: After fixing the comparison and expectations measures, the labs fail to reach the stated revenue threshold by the deadline and the named hyperscalers miss the corresponding expectations. Establishing that the revenue level was necessary also requires a financing and compute-cost analysis.
Counts as contradicted if: The labs remain below the agreed revenue threshold through the deadline while the named hyperscalers meet the relevant expectations, or the labs demonstrably finance and fulfill the expected compute spending without the claimed revenue scale.
“And now the rest of the world has caught up to what I’ve been saying since October 2025 — that hyperscalers need at least $2 trillion in annual AI revenue by 2030 or they’ve wasted their capex.”
Only if: if hyperscalers fail to reach $2 trillion annual AI revenue by 2030
Our reading: Unless hyperscalers reach at least $2 trillion in annual AI revenue by 2030, their AI capex will have been wasted.
Counts as supported if: If combined annual AI revenue fails to reach $2 trillion by the deadline, records must also show that the relevant AI capital investment failed to recover its costs. A revenue shortfall alone is insufficient.
Counts as contradicted if: If the revenue antecedent is met, reliable financial records show that the relevant AI capital investment recovered its costs despite annual AI revenue remaining below $2 trillion.
“If they fail to do so, everybody suffers. Hyperscalers miss revenue estimates. AI data center debt goes unpaid. Venture capitalists find their holdings washed out.”
Only if: if Anthropic and OpenAI fail to reach the revenue their valuations are based on
Counts as supported if: After the specified revenue failure is established, evidence shows all three material consequences: hyperscaler revenue misses, AI data center debt nonpayment and severe venture investment losses, linked to the labs' shortfall.
Counts as contradicted if: After the antecedent is established and sufficient follow-up is available, comparable evidence shows the asserted consequences do not occur. If only some consequences occur, evaluate the components separately rather than treating the entire forecast as supported.
“It involves CoreWeave, IREN, Nebius, and every other neocloud running out of money, fucking over investors in both their stock and debt some time in the next few years, with the underlying collateral made up of otherwise-useless data center construction and GPUs that will, at that point, be in a supply glut rivaling the Atari video game burial.”
Only if: most data center debt goes unpaid, OpenAI and Anthropic fail their obligations, and AI GPU sales grind to a halt
Counts as supported if: The stated collapse occurs within the horizon, and records establish cash exhaustion or insolvency across the claimed neocloud population, losses to both investor classes, and severely impaired collateral amid a broad GPU glut.
Counts as contradicted if: After the antecedent and by the interpreted deadline, at least one company within the universal scope remains able to meet its obligations without the predicted failure, or decisive recovery and inventory records contradict the claimed collateral collapse.
“Analyst expectations have OpenAI and Anthropic contributing at least $444 billion in revenue across hyperscalers in the next three years, and if this revenue fails to arrive — either through insolvency or renegotiation of terms — every connected hyperscaler will see massive revenue misses.”
Only if: if OpenAI/Anthropic revenue to hyperscalers fails to arrive through insolvency or renegotiation
Counts as supported if: After the stated payment failure occurs, every connected hyperscaler reports a large revenue shortfall against the prior analyst baseline, attributable to the lost OpenAI or Anthropic business.
Counts as contradicted if: After a qualifying payment failure and completion of the affected reporting periods, at least one connected hyperscaler avoids a large revenue miss against the prior baseline.
Open-ended windows (9)
These give no end date ("soon", "for years", "at least a decade"), so they have no checkpoint on a calendar.
“At least for another decade or two and probably more, big projects are always going to require software architects who know what they are doing.”
Counts as supported if: Evidence across substantial software projects through the stated duration shows that competent human architectural input remains necessary. Present-day examples alone cannot establish the forecast.
Counts as contradicted if: A well-documented substantial software project succeeds without competent human architectural input during the claimed interval. A qualifying example before 2035-03-02 contradicts both stated durations; an example between 2035 and 2045 depends on the duration interpretation.
“AI might replace average people in some tasks soon; replacing the best humans in a few years? Not going to happen.”
Counts as supported if: At the deadline, credible assessments of frontier capabilities and actual deployments establish that AI still cannot reliably substitute for leading scientists in sustained original research.
Counts as contradicted if: By the deadline, independently verified deployment or a sustained, reproducible demonstration establishes that AI can autonomously substitute for leading scientists in substantive original research.
“At least for the next decade, LLMs (with and without inference time “reasoning”) will continue have their uses, especially for coding and brainstorming and writing.”
Only if: especially coding, brainstorming, writing
Our reading: For at least the next decade LLMs will continue to have uses, especially for coding, brainstorming and writing.
Counts as supported if: Evidence through June 7, 2035 establishes persistent practical usefulness of LLMs, including the emphasized application areas and both reasoning and non-reasoning approaches.
Counts as contradicted if: Evidence establishes that useful LLM applications have ceased before the promised minimum duration, rather than merely declining in popularity or moving to new products.
“Instead, part of the actual outcome over the next few years will be that both countries build a lot of generative AI infrastructure — quite possibly significantly more than they actually need.”
Only if: quite possibly significantly more than needed
Counts as supported if: By the end of the horizon, public records such as company capex filings, official statistics, or data-centre construction and commissioning trackers show that both the US and China built large amounts of generative AI infrastructure during the few years after December 2025.
Counts as contradicted if: By the end of the horizon, records show that either the US or China built only modest new generative AI infrastructure during the period, for example after a sharp, sustained collapse in AI data-centre buildout.
“Then there’s the fact that Apple is suing OpenAI after poaching multiple engineers for its hardware efforts and allegedly both coaching and coercing them into stealing trade secrets, which is all but certain to destroy any chance of OpenAI releasing a device in the next few years…and potentially the company itself.”
Only if: as a result of Apple's lawsuit
Counts as supported if: No qualifying consumer device is released through the interpreted horizon, and reliable records establish that Apple's lawsuit prevented release.
Counts as contradicted if: OpenAI commercially releases a qualifying consumer device within the horizon, or definitive records establish that the lawsuit did not prevent release and any nonrelease had another cause.
“The idea that robots will be the best surgeons on earth anytime soon is absurd, and I challenged him about it on X:”
Only if: anytime soon
Our reading: Robots will not be the best surgeons on earth anytime soon; the idea is absurd.
Counts as supported if: At the deadline, sufficiently comprehensive clinical and regulatory evidence shows that robots have not attained the specified superiority.
Counts as contradicted if: By the deadline, credible comparative clinical evidence establishes autonomous robotic superiority to leading human surgeons across a representative range of surgery.
“Neither of these companies can actually afford this compute, but that doesn’t matter, because this data center won’t actually exist for years, if it ever does.”
Our reading: The data center under the Anthropic-Lambda deal will not exist for years, if ever.
Counts as supported if: Reliable records establish that the identified facility did not begin delivering the contracted compute service before the two-year anniversary of publication.
Counts as contradicted if: Reliable commissioning and customer-service records establish delivery of the contracted compute service before the two-year anniversary.
“In any case, while we don’t know much about Anthropic past Q3 2026, we do know that it has hundreds of billions of compute costs due in the next few years, and that they’re going to increase dramatically:”
Our reading: Anthropic's compute costs, hundreds of billions due over the next few years, will increase dramatically.
Counts as supported if: Payment schedules establish at least $200 billion due within the interpreted horizon, and comparable annual results show substantially higher compute costs than in 2026.
Counts as contradicted if: Complete records for the interpreted horizon establish less than $200 billion due, or comparable annual compute costs remain flat, decline, or increase only slightly.
“In fact, the literal opposite is going to take place over the next few quarters, as Anthropic’s massive non-cancellable burdens start weighing on its balance sheet regardless of how well the company does.”
Only if: regardless of how well the company does
Counts as supported if: Results within the interpreted horizon show operating losses or material deterioration from Q2 profitability, with disclosures attributing a material part of that pressure to payments or expenses under the non-cancellable compute obligations.
Counts as contradicted if: Across the interpreted horizon, comparable operating profitability is maintained or improves while the relevant commitments take effect, without the predicted financial strain. Strong revenue alone is insufficient.