Reference Log

Reported The buyback squeeze, restated on verified figures. Alphabet repurchased zero stock in Q1 2026 — its first zero quarter in a decade — and zero again in Q2, against $13.24B in the year-ago quarter and $28.31B across the comparable half. That halts a programme that returned close to $300B over five years. Across the largest technology firms, buybacks have fallen roughly 17% over the past year while repurchases across the rest of the S&P 500 continued to rise. Alphabet has meanwhile raised roughly $70B in combined equity and debt and doubled long-term debt to $98.2B.

Reported Dot-com earnings damage was worse than this monitor previously stated. S&P 500 trailing operating EPS fell by more than half from 2000 to 2002, against a 1960–2000 average recessionary decline of about 13%. The 2001 drop was extreme relative to all prior recessions specifically because tech-bubble and Y2K capital-spending excesses reversed. The index bottomed at 776 on 10 October 2002, 49% below the March 2000 peak.

Filing Neo-cloud GPU useful lives are not uniform. CoreWeave depreciates technology equipment including GPUs straight-line over six years (disclosed in its S-1). Nebius uses four years, closer to Nvidia's roughly three-year architecture cadence. The neo-cloud model is a financed wager on whether the useful economic life of a GPU is 5–6 years or 2–3.

Reported Grid interconnection lead times, restated as a range. Securing grid power for a new data centre in 2026 typically takes 24 to 72 months, with large loads in constrained regions quoted at five to seven years. Substation transformer lead times now exceed 160 weeks. ERCOT was tracking a large-load interconnection queue of approximately 410 GW as of April 2026, roughly 87% of it data centres; the national queue backlog is about 2,600 GW, and only about a quarter of active queue capacity holds an executed or draft interconnection agreement.

Inference Monitor-computed figures, stated as such. Two numbers on the dashboard are this monitor's own computations from filed inputs rather than reported figures, and cannot be externally corroborated: the CIP-to-net-PP&E ratio of 28.4% across the top-five cohort, and Microsoft's cumulative incremental ROIC of 24.5% (AI era) against 51.2% (pre-AI). The CIP ratio derives from segment profit and asset disclosures; the iROIC pair is computed on a consolidated basis, because segment invested capital is not disclosed and a segment-level figure would rest on an allocation assumption the filings do not supply. They are carried because the method is consistent quarter to quarter, not because a third party publishes them. Note further that since August 2026 the CIP ratio reads a floor rather than a measurement, see [R18].

Inference All four hyperscalers earn a lower pretax return on invested capital than at the start of 2024, and the "air gap" does not explain it. Pretax ROIC computed quarterly from SEC XBRL as (quarterly operating income × 4) ÷ invested capital, where invested capital is stockholders' equity plus non-current long-term debt plus non-current operating lease liabilities, less cash and short-term investments. Operating income grew 36% to 79% over the period while invested capital grew 67% to 160% (Microsoft +67%, Meta +120%, Amazon +149%, Alphabet +160%), so the fall is denominator-driven, not an earnings failure. Removing assets under construction from the denominator — capital that is in the base but not yet earning — lifts the level by roughly 10–12 points but steepens rather than flattens the decline: Alphabet −6.9pp becomes −7.5pp, Meta −14.6pp becomes −15.7pp. Q4 is absent by construction because 10-Ks report full-year durations; Microsoft's fiscal year ends in June.

Inference On measured cost per task, moving from a 70.8% to a 72.9% benchmark score costs 6.4× — $2.81 to $18.02 per task — and one model's reasoning-effort setting moves cost 2.6× on its own. CursorBench runs ambiguous, multi-file coding tasks drawn from real sessions and reports, per model configuration, the score, the cost per task, the tokens per task and the agent steps per task. Cursor states cost is computed by applying each model's published per-million-token pricing for input, cache read, cache write and output to the tokens it used, so the dollar figure carries none of this monitor's own pricing assumptions — no blended rate, no hosted-price choice. 67 configurations; 11 on the cost frontier. This independently reproduces the shape in [R76], which reached the same conclusion from list prices and a capability index. Scope is narrow: one benchmark, agentic software work only, and the 6.4× step is a jump between two specific models rather than a smooth market curve.

Inference A single agentic coding task consumes at least 1.5 million tokens, and the cost of such work is driven by token volume rather than by the per-token rate. Solving implied = R × price_in + price_out against vendor input and output prices across 32 configurations gives an input:output ratio of 16× to 50× (median 32×), consistent with agents resending accumulated context across 17 to 99 steps. Reconstructed, Claude Fable 5 at maximum effort spends roughly 1.55M tokens per task at an effective $11.65 per million, against the $20.00 blended list rate this monitor's other panels use — so per token, agentic work is cheaper than assumed, not dearer. THIS IS A BOUND, NOT AN ESTIMATE. Cursor does not define what its "tokens per task" column comprises; the arithmetic implies output only. Because cache reads bill at roughly a tenth of the input rate, the reconstruction overcharges those tokens, so 1.5M is a floor and $11.65 a ceiling — both corrections push the same way, which is why the conclusion survives the undocumented field. The consequence for [R76] and [R78]: no fixed factor converts a per-token axis into cost per task in either direction, which is why those panels state what they measure rather than adjusting for it.

Inference Million-token tasks are not specific to agentic coding. Effort-matched, an agentic coding task costs a single-digit multiple of an abstract-reasoning task on the identical model configuration: median 1.72× against ARC-AGI-2 and 4.20× against ARC-AGI-1. CursorBench (agentic software work) and ARC-AGI-2 (abstract visual reasoning — no repository, no tool loop, no resent context) both publish cost per task and both label the exact model configuration including its effort setting. For a configuration appearing on both, the price per token is the same number on both sides and cancels exactly, so the ratio of costs is the ratio of token volume. This uses no price data, no input:output blend, and does not depend on this monitor's price file being correct — a stronger instrument than the reconstruction in [R81]. Across 32 such configurations the ratio runs 0.43× to 4.15×, median 1.72×, and the spread is a model-family property rather than noise: all five GPT-5.6 Luna configurations are cheaper on coding, all five GPT-5.6 Sol configurations are 3× or more the other way. Effort-exact matching is required, not a refinement — the effort dial is the cost dial ([R81]: 2.6× within one model), and the two benchmarks test different effort mixes, so matching on the base model alone compares a high-effort run against a low-effort one. The point estimate is not stable and must be quoted as a range. ARC-AGI-2 was the benchmark selected; the adjacent one, ARC-AGI-1, answers 2.4× differently across 33 paired configurations. The mechanism is difficulty rather than a defect — ARC-AGI-1 is far easier (median score 0.526 against 0.053), so models spend fewer tokens on it and the ratio rises. The order of magnitude replicates on both benchmarks; the single figure does not. Scope: all three benchmarks with measured per-task cost are long-generation work, and two of the three are the same domain. Chat, summarisation, classification and retrieval remain untested — no single-turn benchmark publishing cost per task was available.

Withdrawn Withdrawn 21 August 2026: "nothing here generalises to chat, summarisation, classification or retrieval." Published alongside [R81] as the scope caveat on cost per task. It was honest but untested, and [R82] tested it: on a second domain with a method that needs no pricing assumption, the factor is a median 1.72× rather than the order of magnitude the wording implied. The claim is replaced by the narrower one in [R82] — two domains tested and held, the remaining single-turn domains still untested.

Inference Crossing from half-hour to hour-long autonomous tasks costs 19.2×; going from a one-hour to an eight-hour horizon costs about 5× in total. The cheapest model able to sustain a task of each length, against blended list price per million tokens (3:1 input:output), on the same price basis as [R76]. Capability is METR's task-completion time horizon — the task duration at which a model succeeds 50% of the time. 39 of 43 models with a published horizon carry a price. This is the opposite shape to the capability-index frontier in [R76], where the steep step sits at the top of the range. Both hold; they are different axes, and this one answers the question a buyer faces. Three of the six models on this frontier are no longer sold — OpenAI has withdrawn o4-mini and GPT-5 from its price list, Moonshot has withdrawn K2 Thinking — so those figures are historical. That is structural: measuring long-horizon capability takes long enough that the models measured have since been retired. Fitted intervals are wide (Claude Opus 4.6: 5.3h to 65.8h), so the shape is defensible and the top-end level is not. The cost axis is price per token, not per task, and a multi-hour agentic job emits far more tokens than a short answer, so the real spread is wider than shown.

Inference Measured at an 80% success threshold rather than 50%, task horizons collapse by roughly an order of magnitude, and the most expensive model on the frontier is beaten by one costing less than half as much. Claude Opus 4.6 falls from a 12.0h horizon at 50% to 1.17h at 80%; Gemini 3.1 Pro falls from 6.4h to 1.50h. At the stricter bar Gemini sustains the longer horizon at $4.50 against Opus 4.6's $10.00, so Opus 4.6 leaves the frontier entirely. Both thresholds are METR's own published figures; 14 of the priced models carry both. A 50% success rate is a coin flip, not a capability anyone deploys against. The buildout is underwritten by an expectation of long autonomous task capability; at a threshold an operator would rely on, the longest horizon in this dataset is about ninety minutes, and paying twice as much does not extend it. Same caveats as [R78] — fitted intervals are wide, the newest and most expensive models carry no horizon at all, and the cost axis is per token rather than per task.

Inference On the capability/cost frontier, each quarter of the capability range costs more than the one before it — 1.9×, 3.5×, 4.5× — and the 75th-to-90th percentile band alone costs 10×; a 3.4% gain in measured capability at the top costs 22×. The frontier is the Pareto set — the cheapest model available at or above each capability level. Capability is the Epoch Capabilities Index; cost is list price per million tokens blended 3:1 input:output. 200 of 283 indexed models carry a published price. Every unpriced model in the top forty by index was priced by hand from its vendor's page, because an unpriced model is silently dropped from the Pareto set and dropping a cheap capable one would exaggerate the premium. Two so priced — Kimi K3 and Grok 4.20 — do not reach the frontier and are recorded as negative results. Composition splits at index 156: five of the six frontier models below it are open-weight or Chinese, none of the five above it are. Price basis differs by model type: the creator's list price where the creator sells the model, the cheapest third-party hosted rate where it only released weights (hosted rates for one model span up to 19× across providers). The index does not cover every lab — Tencent has no entry, so its models cannot appear at any price, which flatters the premium rather than understating it. The cost axis is price per token, not per task — a verbose reasoning model emits more tokens for the same question, so this understates its true cost per task, and understates it most at the top of the range.

Reported Gemini 3.7 Flash's $0.75 / $3.75 per million tokens is promotional, stated to hold through 31 December 2026 and to double to $1.50 / $7.50 on 1 January 2027. Read directly from Google's published pricing page, which prints the expiry alongside the rate. At the post-expiry rate the model's blended price moves from $1.50 to $3.00 and it leaves the capability/cost frontier, since GPT-5.6 Terra offers materially more capability at $4.50. An aggregated price feed reports the current number with no expiry attached, so this was visible only by reading the vendor's own page — the same class of trap as reading commitment totals from structured XBRL rather than filing text.

Inference Operating cash flow as a multiple of capital expenditure has compressed toward 1.0× across every tracked builder since 2022; Oracle and Amazon have crossed below it. Trailing-twelve-month operating cash flow divided by trailing-twelve-month capex, per company per quarter, from cash-flow-statement figures in SEC filings. Below 1.0× the buildout is no longer self-funded from operations and the balance is met by debt, leases or external capital — which is the transition this monitor tracks into the credit leg.

Inference Off-balance-sheet contracted obligations across the big four total ~$2.35tn, against ~$511bn of trailing-twelve-month capex — 4.6× in aggregate. Components, read from filing text: Alphabet $811.0bn purchase commitments plus $85.2bn of leases not yet commenced; Meta $349.3bn non-cancelable commitments plus $279.0bn uncommenced leases; Microsoft $194.1bn purchase commitments plus $34.6bn construction commitments plus $329.1bn uncommenced leases; Amazon $130.1bn unconditional purchase obligations plus $137.2bn uncommenced leases. TTM capex: Alphabet $132.4bn, Meta $89.3bn, Microsoft $115.9bn, Amazon $173.0bn. Coverage -- years of current capex already contracted -- is therefore Alphabet 6.8×, Meta 7.0×, Microsoft 4.8×, Amazon 1.5×. On-balance-sheet lease liabilities are excluded for all four so the cohort sits on one basis. Microsoft's $329.1bn is disclosed directly in the FY2026 10-K lease note — “additional leases, primarily for datacenters, that had not yet commenced of $329.1 billion”, commencing between fiscal 2027 and fiscal 2033 with terms of 1 to 20 years. It also reconciles: $443.5bn of total lease obligations less $114.4bn of commenced payments due. This is the largest single long-dated item in the cohort and the filing attributes it to datacenters. Alphabet does not quantify the compute-versus-energy split. Its Note 10 reports $707.0bn of commitments on contracts with remaining terms over one year and says only that “the significant majority” relates to long-term supply agreements for technical infrastructure; energy service agreements are the minority but carry much the longer tail — terms of 2 to 26 years running to 2054, with take-or-pay minimums and substantive termination fees, against supply and content commitments expected to be fulfilled generally through 2030. Dates differ by filer — Q2 2026 for Alphabet, Meta and Amazon; FY2026 ending 30 June for Microsoft — and Microsoft's step is therefore year-on-year where the others are quarter-on-quarter. These figures are not tagged in XBRL; the structured-data endpoints return narrow, unrelated items (Alphabet's tagged PurchaseObligation reads $7.7bn against $811.0bn in the filing text), so anyone reproducing this from an API rather than the filings will conclude the reporting is wrong.

Withdrawn "Established data centre operators generate 5–8% pre-tax returns on capital." This figure is carried in this monitor's supporting framework analysis (it does not appear on the dashboard) and could not be sourced on re-verification. Publicly available material on Equinix and Digital Realty reports growth, market capitalisation and dividend metrics rather than pre-tax return on capital. The underlying claim — that colocation operators earn low single-to-mid-digit capital returns while trading at technology multiples — may well be right, but it is not currently evidenced here and has been removed from the dashboard pending a primary source.