brain/
questionopenstock-market

Is 2026 AI inference revenue ~$70-80B (Jain) or well over $200B (Baker) — and is xAI's Colossus running at 11% utilization or a 55% ARR?

Notes

Is 2026 AI inference revenue ~$70-80B (Jain) or well over $200B (Baker) — and is xAI's Colossus running at 11% utilization or a 55% ARR?

The dispute

Two professional investors, speaking three days apart in June 2026, put the same numbers an order of magnitude apart. Both are cited in this wiki. Neither is obviously wrong on its face, and the gap is large enough to invert the AI-capex conclusion.

The bear side — rajiv-jain (GQG Partners, $160B AUM), 2026-06-08.

  • On revenue: "The cumulative capex of all these mag companies in their history is one and a half trillion dollars. Think about it now they're talking about three trillion in just three years... you're spending a trillion dollars a year and the revenue on AI talk about maybe 70, 80 billion." (From 2026-06-08-podcast-capital-allocators-contrarian-quality-at-gqg-partners-rajiv-jain-ep.)
  • On Colossus: xAI (not OpenAI) — "x AI cash losses were give and take double digit billions 12 to 15 and their capacity utilization on the Colossus was 11% now, now they're selling the capacity to anthropic." The 2026-07-09 ingest mis-attributed Colossus to OpenAI; the transcript attaches "their" to xAI. Cash losses "12 to 15" billion.
  • Conclusion: "Our view is that this is a powerful technology but the economics are really bad and time is not a friend."

The bull side — gavin-baker (Atreides Management), 2026-06-11.

  • On revenue: gavin-baker in 2026-06-11-podcast-bg2-pod-the-spacex-ipo-fable-5-ai-capex-update-market: "I think we end this year well over 200 billion in inference revenue. Well over."
  • Against the $300B consensus figure being an overstatement, he argues the opposite: "I think that 300 billion is low, man"
  • On Colossus: "Your colleague at Altimeter, Freda, also, she calculated a 55% ARR on Colossus 1. If you can borrow money at 6, 7, 8% and invest in something with a 55% arrow"
  • On the fraction of capex that is not revenue-generating: "so we'll call it 35% is spending that's not revenue generating that is going to kind of make the next model."

The framing question, from brad-gerstner in the same source: "So we're spending 1.5 trillion of capex on 300 billion of inference revenue. Does that math math for you?"

Why it matters

The AI-capex cluster is the largest single vertical in this project's signal feed. hbm-cowos-as-binding-bottleneck (active, high conviction), ai-capex-to-power-and-materials-cascade, and inference-demand-to-wafer-scale-advantage all rest on capex being sustained. Jain's figures imply the spend is unsupported by revenue and therefore fragile; Baker's imply it is comfortably covered and accelerating. The same Colossus asset is cited by both — as evidence of stranded capacity and as evidence of a 55% return.

compute-utilization-overhang-as-latent-supply currently encodes the low-utilization reading as a bound on the scarcity thesis. Baker's ARR figure, if right, removes that bound.

What would resolve it

  1. Definitional reconciliation. Jain says "revenue on AI"; Baker says "inference revenue." These may not be the same quantity — one may be lab revenue net of training, the other gross inference billings including internal transfer. Neither speaker defines the term. This is the most likely explanation and the cheapest to check.
  2. Whose Colossus? Resolved 2026-08-17. The original Jain transcript attributes Colossus to xAI, not OpenAI; the 2026-07-09 "OpenAI's Colossus" line was a wiki mis-record. Baker's ARR is on "Colossus 1" specifically.
  3. Utilization vs. return are not the same metric. An 11%-utilized asset can still post a high ARR if the sold fraction is priced high enough. The two claims may both be true and non-contradictory — which would itself be the resolution.
  4. A disclosed segment number from any hyperscaler that separates inference billings from training spend.

Status

Adjudicated 2026-07-09 during the Bg2 ingest: record both, move no conviction. Neither source is downgraded; no mechanism's conviction was changed on the basis of this dispute. Both speakers are talking their book in opposite directions — Jain is short/underweight the complex, Baker is long it and a SpaceX shareholder.

Narrowed 2026-08-17 via 2026-08-17-autoresearch-jain-baker-ai-inference-revenue. Still status: open — there is still no official 2026 "inference revenue" line. The four resolvers moved as follows. No mechanism conviction was changed.

  1. Definitional — narrowed, most likely the gap. Jain's June "$70, 80 billion" matches a then-current frontier-lab stack (OpenAI ~$20B ARR at start-2026 + Anthropic's May $47B run-rate ≈ $67–72B). Baker's "well over $200B" is an end-2026 forecast of a broader inference bucket, the same stack he sketched on 2026-05-22 (OpenAI + Anthropic + Gemini + Cursor + xAI + open source). Syndicated 2026 "inference market" prints fetched in the same pass range from ~$9B (inference-platform) to $23B (inference-as-a-service) to $118B (hardware-inclusive TAM) — none is Baker's number, and none is a 10-K line. From 2026-08-17-autoresearch-jain-baker-ai-inference-revenue.
  2. Whose Colossus? — resolved. The original Jain transcript attributes Colossus to xAI. The 2026-07-09 "OpenAI's Colossus" line was a wiki mis-record. Grokipedia's Colossus page (cited in the same source) is xAI/Memphis.
  3. Utilization vs return — narrowed to both-can-be-true. The 11% figure, as reported by The Information on 2026-05-02 and restated as MFU (model FLOP utilization) by later write-ups and by anjney-midha on 2026-06-13, is training throughput over theoretical peak FLOPs, not occupancy of sold capacity. Baker's 55% is a financial return on Colossus 1 as a rented asset (vault transcript: ARR; DruckFin recap: IRR). Selling the cluster to Anthropic — which Jain himself notes — is how a low-MFU training fleet still posts rent. From 2026-08-17-autoresearch-jain-baker-ai-inference-revenue.
  4. Hyperscaler inference-vs-training split — still open. Closest primaries: Microsoft FY2026 10-K related-party revenue from OpenAI of $24.1B (no compute / revenue-share / inference / training cut); satya-nadella "AI business" $37B ARR as of Q3 FY2026 (no product-line cut); Alphabet Q2 2026 Cloud $24.768B (+82%) naming "enterprise AI Solutions and enterprise AI Infrastructure" without splitting them. Adding Microsoft's $24.1B to OpenAI ARR double-counts. From 2026-08-17-autoresearch-jain-baker-ai-inference-revenue.

August 2026 lab floor (does not close the question): sarah-friar disclosed OpenAI $40B ARR (2026-08-14) with enterprise now larger than consumer; Anthropic preliminary Q2 revenue >$11.5B (vs $4.73B Q1; May run-rate $47B). Two-lab annualized floor ~$86B–$100B — past Jain's June snapshot, still less than half of Baker's year-end $200B+ unless hyperscaler AI ARR is folded in. ARR ≠ booked revenue (OpenAI FY2025 booked $13.07B). From 2026-08-17-autoresearch-jain-baker-ai-inference-revenue.

Evidence we have

Sources

Related

Referenced by
Mechanisms
Frontier-lab vertical integration → enterprise trust break → open-model sovereign stack (Nvidia + Palantir rerate, model-layer margin pressure)IBM pre-announcement → enterprise capex reallocates to supply-constrained infrastructure → SaaS derates while hardware + security bid
Concepts
Compute utilization overhang as latent supplyOpen-source share shift is bullish, not bearish, for computeOperating profit per gigawattToken-price inflation favors asset-heavy compute owners
Entities
AnthropicBrad GerstnerDavid SacksGavin BakerOpenAIRajiv JainSarah FriarSatya Nadella