On the Big Technology Podcast (August 12, 2026), Paul Kedrosky argued that an 80 percent compounding annual token price decline requires roughly 100 million-fold token volume growth over six years to keep the buildout's revenue engine intact, making the AI data center buildout unjustifiable. The math does not check out. An 80 percent annual decline for six years cuts prices to 1/15,625 of their starting value: flat revenue needs about 15,600x token growth, not 100 million. The 100 million figure requires a 95 percent annual decline, which only appears in fixed-capability price series - the cost of replicating a specific historical model's benchmark score - not in what buyers actually pay. Frontier flagship prices fell 35-45 percent per year over this period, and OpenAI's flagship output price rose from $10 to $30 per million tokens between late 2024 and mid-2026. The realized blended revenue per token declined roughly 65-75 percent annually. Kedrosky's specific claim is wrong by four orders of magnitude against his own stated premise.
The underlying concern, though, deserves to be taken seriously. Under a realistic 65-75 percent annual price decline, justifying Sequoia's $3 trillion revenue target requires 5.6-7.1x annual token volume growth, sustained for six years. The industry currently delivers 5-10x per year, so the threshold is being met today. The question is whether it holds. Google's own token volume growth decelerated from 50x to 7x year over year. Sustaining the required pace for six years without meaningful deceleration is the actual open question, and it is a hard one.
The capex-revenue gap is real regardless of which arithmetic you use. The four big hyperscalers have deployed roughly $1.05 trillion in capital since 2023, and AI-specific revenue runs at $100-135 billion annually today. The 100 million-fold framing overstates the case so severely that it invites dismissal. The version of the argument that survives scrutiny is narrower, more conditional, and still worth taking seriously.
Kedrosky's chain: token prices have fallen 70-80%/yr on a constant-performance basis for four years; data centers are financed like commercial real estate expecting a ~6-6.8% cap-rate return; therefore Jevons-paradox rescues require token volume to grow "around 100 million-fold" over six years, which he calls possible but not likely.
| Scenario (6-year horizon) | Annual price decline | Token growth needed for flat revenue |
|---|---|---|
| His stated premise | 80% | 15,625x |
| His lower bound | 70% | 1,372x |
| What his 100M-fold figure actually implies | 95.4% | 100,000,000x |
| Frontier list price (measured, OpenAI flagship) | 35–45% | ~21x |
| Blended realized $/token (measured, triangulated) | 65–75% | 1,372–4,096x |
The whole dispute reduces to which price series you divide revenue by. The three series behave completely differently.
| Series | Measured decline | What it describes |
|---|---|---|
| Fixed capability (cost to hit an old benchmark bar) | 90–99.5%/yr | Epoch AI: 9x-900x/yr across benchmarks, median ~50x/yr. The "2010 CPU" series: real, and irrelevant to revenue, because nobody buys the old bar. |
| Frontier list price (what the current flagship costs) | ~35–45%/yr | GPT-4 ($30/$60 in Mar 2023) to GPT-5.6 Sol ($5/$30 in Jul 2026). Claude Opus held at $5/$25 across five straight releases, and Anthropic added a premium tier above it ($10/$50). OpenAI flagship output price rose from $10 (GPT-4o) to $30 (GPT-5.5/5.6). |
| Blended realized $/token (revenue ÷ tokens actually served) | ~65–75%/yr | Triangulated from OpenAI disclosures: tokens up ~10.5x in 13 months while revenue grew ~2.5-3x. Confounded by flat-fee subscribers consuming ever more tokens. Not disclosed by any provider; order-of-magnitude only. |
The counter-argument that frontier prices stay high is half-right. Frontier list prices are indeed the slowest-declining series, and users do migrate up-generation continuously, exactly like CPU buyers. But the revenue-relevant series is the blended one, and it falls roughly twice as fast as frontier list prices because of mix shift toward cheap models (open-weight models went from ~1% to ~33% of OpenRouter volume in a year), aggressive budget-tier cuts (GPT-5.6 Luna, minus 80% in one move in July 2026), and flat-fee subscriptions serving more tokens for the same dollar. The saving grace is token-per-task inflation running the other way: reasoning models consume 15-20x the tokens of non-reasoning models on the same task, reasoning share went from negligible to over half of tokens in a year, and average prompt length quadrupled. Falling unit prices and inflating token intensity largely cancel: that is why OpenAI and Anthropic gross margins are flat-to-slowly-improving (33-44%) rather than exploding in either direction.
Flat revenue is the wrong bar: the buildout needs revenue to grow into the capex. Using Sequoia's July 2026 framing ($1.5T/yr of AI infrastructure spend needs ~$3T/yr of AI revenue) against today's $100-135B run-rate, revenue must grow 22-30x. Compounding that with the measured 65-75%/yr realized price decline:
Token counts overstate demand: Google's 3.2 quadrillion monthly tokens include Search AI Overviews and free-tier usage that monetize at zero. The cleaner demand signal is paid cloud revenue, and it tells a different story than the token series - accelerating, not decelerating.
| Provider | Q1 2023 growth | Q2 2026 growth | Note |
|---|---|---|---|
| Google Cloud | +28% | +82% | $24.8B quarter; operating margin went from single digits to ~36%. While Google's token growth decelerated 50x→7x, its cloud revenue growth tripled: proof the two series measure different things. |
| AWS | +16% | +37% | $42.2B quarter, fastest in 18 quarters, five straight quarters of acceleration. AI business and custom-chip business each crossed $25B annualized, each more than doubling YoY. |
| Azure | +27% | +43% | Crossed $100B annual revenue. AI services were adding 8→13 points of Azure growth through 2024 and grew 157% YoY as of late 2024. |
| Oracle OCI | n/a | +93% | $5.8B quarter (calendar Q2 2026). |
| CoreWeave | — | ~+112% | $2.58B quarter; $104B backlog, +246% YoY. |
| Nebius | — | +684% (Q1) | Guiding ~540% ARR growth in 2026. |
Eighteen distinct claims were extracted from the episode. The load-bearing ones, tested against primary data:
| Claim | Status | Evidence |
|---|---|---|
| Capex trajectory: ~$350-400B → ~$700B (2026) → ~$1.5T (2027) | Supported | Big-4 actuals: $228B (2024), $376B (2025); 2026 guidance sums to ~$750B for the Big 4 ($895-970B with Oracle and CoreWeave). Morgan Stanley projects $800B (2026), $1.16T (2027) globally. His 2027 number is the high end of street estimates. |
| Token prices fall 70-80%/yr "constant performance basis" | Partly | Understated for fixed capability (literature says 90-99%/yr), overstated for what buyers pay (frontier 35-45%/yr, blended ~65-75%/yr). The phrase "constant performance" is doing unacknowledged work. |
| Offsetting the decline needs ~100M-fold token growth in 6 years | Contradicted | His own premise implies 1,400-15,600x. 100M-fold requires a 95%/yr decline that exists only in the fixed-capability series. Off by ~4 orders of magnitude. |
| External financing now >50% of data-center funding (Q2 2026) | Plausible, unverified | Directionally consistent with Morgan Stanley's $1.5T financing gap (Big 4 self-fund ~$1.4T of $2.9T through 2028), record tech IG/HY issuance, and NVIDIA's $500B asset-manager financing platforms (Aug 10). The specific >50% Q2-2026 figure could not be independently confirmed. |
| GPU economic life shorter than hyperscaler 5-6yr accounting | Supported as live dispute | Amazon shortened a subset of servers 6→5yrs citing AI hardware pace, while Meta extended; neoclouds use 4-5yrs; Burry argues 2-3yrs (~$176B understated depreciation 2026-28). The hyperscalers themselves now disagree, which is his point. |
| Buildout exceeds railroads, electrification, interstates as share of economy | Contradicted for railroads | AI capex is ~1.3-1.5% of GDP (Epoch AI), above the telecom-2000 peak (~1.2%) but far below peak railroad investment (6-20% of GDP in mania years). Correct vs. fiber/telecom; wrong vs. the 19th century. |
| Demand can't rescue revenue (Jevons "innumeracy") | Open | Required ~6x/yr; delivered 5-10x/yr and at threshold. Rescue is neither implausible (his framing) nor assured (the bulls'). Depends entirely on whether growth sustains or decays. |
| Model convergence pushes competition to price | Supported | Open-weight share of OpenRouter volume ~1%→~33% in a year; budget tiers cut 80% while flagships hold; Chinese models used for cost-sensitive workloads. |
| Losses would metastasize via credit markets (GFC analogy) | Structurally supported | Tech now the largest slice of HY issuance; GPU-collateralized structures (NVIDIA/Apollo/Blackstone/BlackRock/Brookfield/GS/KKR) explicitly modeled on infrastructure securitization. Whether that IS 2008 depends on leverage ratios that are still low (hyperscaler D/E ~0.23 vs telecom's). |
1. Buildout scale exceeds all prior US infrastructure cycles (00:01-00:02). 2. Compressed timeline removes market stop-points (00:04). 3. External financing >50% as of Q2 2026 (00:03). 4. Capex accelerating $350-400B→$700B→$1.5T (00:05). 5. CRE/cap-rate analogy flawed, duration mismatch (00:07-00:13). 6. GPU replacement driven by MTBF vs generation vs architecture; training chips wear like race cars, inference like church cars; some failing at 18 months (00:11, 00:17). 7. Tokens are hyperdeflationary unlike rail/electricity/fiber (00:12). 8. Jevons rescue needs ~100M-fold growth (00:19-00:21). 9. Model convergence, Pepsi/Coke blind tests (00:19, 00:31). 10. Labs going up-market into apps is evidence they see the collapse coming; if the tech generated reliable alpha they'd keep it (00:21, 00:28). 11. Training capex increasingly wasted; gains now from harnesses and post-training (00:55). 12. Losses metastasize like 2008 via insurance/private credit (00:57). 13. Check-size filter distorts allocation; sovereigns can only write $100B checks (00:42-00:44). 14. First bubble combining tech + credit + policy + real estate (00:45). 15. "This time is different" reflexivity is itself unprecedented (00:35). 16. China more insulated; provincial overbuild pattern (01:02). 17. His steelman: the call-option-on-AGI argument is undiscountable; he refuses the frame but concedes it is the one path where the buildout is justified (00:38, 00:53). 18. The ending is overdetermined: rates, IPO disappointment, export controls, government stakes, or training-capex cuts could each break it; no date given, and he did not reject the host's "year and a half" framing (00:51, 01:01). He also disclosed working with two hedge funds on positions against the cycle for close to a year.
| Quarter | Microsoft | Alphabet | Amazon | Meta |
|---|---|---|---|---|
| Q1 2023 | 6,607 | 6,289 | 14,207 | 6,823 |
| Q2 2023 | 8,943 | 6,888 | 11,455 | 6,134 |
| Q3 2023 | 9,917 | 8,055 | 12,479 | 6,496 |
| Q4 2023 | 9,735 | 11,019 | 14,588 | 7,592 |
| Q1 2024 | 10,952 | 12,012 | 14,925 | 6,400 |
| Q2 2024 | 13,873 | 13,186 | 17,620 | 8,173 |
| Q3 2024 | 14,923 | 13,061 | 22,620 | 8,258 |
| Q4 2024 | 15,804 | 14,276 | 27,834 | 14,425 |
| Q1 2025 | 16,745 | 17,197 | 25,019 | 12,941 |
| Q2 2025 | 17,079 | 22,446 | 32,183 | 16,538 |
| Q3 2025 | 19,394 | 23,953 | 35,095 | 18,829 |
| Q4 2025 | 29,876 | 27,851 | 39,522 | 21,383 |
| Q1 2026 | 30,876 | 35,674 | 44,203 | 18,997 |
| Q2 2026 | 35,802 | 44,924 | 54,208 | 30,116 |
Microsoft fiscal quarters remapped to calendar. 2026 guidance: Microsoft ~$190B, Alphabet $195-205B, Amazon ~$220B, Meta $130-145B (incl. finance leases), Oracle FY27 ~$70B net, CoreWeave $35-39B.
| Provider | Metric | Points | Growth |
|---|---|---|---|
| Tokens/month, all products | 9.7T (May 2024) → 480T (May 2025) → 3.2Q (May 2026) | 50x, then 7x YoY | |
| API tokens/min | 7B (Oct 2025) → 16B (Apr 2026) → 19B (May 2026) | ~6x since Jan 2025 | |
| OpenAI | API tokens/min | ~300M (Nov 2023) → 6B (Oct 2025) → 15B (Mar 2026) | ~9-10x/yr latest, accelerating |
| OpenAI | Weekly active users | 100M (Nov 2023) → 800M (Oct 2025) → 900M (Feb 2026) | Decelerating (S-curve) |
| Microsoft | Foundry tokens | >100T in Q3 FY25 (5x YoY); >500T FY25 (7x YoY); metric changed FY26 | 5-7x/yr |
| Anthropic | Revenue run-rate (proxy) | $1B (Jan 2025) → $9B (Dec 2025) → $47B (May 2026) | 9x/yr, then accelerating sharply |
| OpenRouter | Mix shifts | Open-weight share ~1%→~33%; reasoning share negligible→>50%; prompts ~1.5k→>6k tokens | Intensity inflating |
Growth is decelerating at Google (50x→7x) and in OpenAI user counts, but accelerating in OpenAI per-user token intensity and Anthropic revenue. Microsoft's switch from disclosing aggregate tokens to counting large customers is itself a signal worth watching.
The blended realized $/token estimate (65-75%/yr decline) is a triangulation from public soundbites, not an audited metric: it divides estimated revenue by disclosed token throughput at two points and is confounded by flat-fee subscription tokens, scope ambiguity between API-only and total tokens, and run-rate vs. booked revenue. No provider discloses this number. The required-growth band (5.6-7.1x/yr) inherits Sequoia's $3T target, which itself embeds a 2x infrastructure-cost gross-up and a 50% end-customer margin assumption; Bain's independent method lands at $2T/yr by 2030. AI-specific revenue ($100-135B) double-counts some Azure-resold OpenAI capacity and excludes Google Cloud's AI share, which is not separable from its $99B annualized total. Kedrosky's episode figures were taken from the auto-generated transcript and could contain transcription errors on numbers; the 100M-fold figure appears clearly and twice, so it is not a transcription artifact.
Podcast: Big Technology Podcast, "Here's How The AI Bubble Bursts," Aug 12, 2026 (transcript via podscripts.co). Pricing: Epoch AI inference price trends · Stanford AI Index 2025 ch. 1 · a16z LLMflation · OpenAI pricing · CNBC on GPT-5.6 cuts · Decoder on Opus 4.5. Usage: Google I/O 2025 · Register on I/O 2026 · OpenAI Mar 2026 · Microsoft FY26 Q4 call · Anthropic run-rate series · OpenRouter State of AI. Capex & break-evens: company 10-Q cash-flow statements via stockanalysis.com · Alphabet Q2 2026 · Morgan Stanley financing gap · Sequoia $600B question · TechCrunch on Cahn's $3T update · Bain $2T · CNBC on GPU depreciation · Epoch AI capex share of GDP · CNBC on NVIDIA financing platforms.