Is There Really a Shortage of Compute?

GPU queues, memory shock, the power wall, the compute divide — and the bubble case — a grounded 2026 verdict

Deep Research report — published September 10, 2026 — grok.gpbot.net

Executive Summary

• Yes — a shortage is real and measurable in 2026, but it is not a single shortage of "GPUs". It is a stack of simultaneous bottlenecks: GPUs (36–52-week lead times), HBM/DRAM memory (spot prices up ~8× since early 2025), TSMC advanced-node capacity (sold out through 2027), and above all electricity — gas turbines sold out through 2029, transformers on multi-year lead times, grid interconnections taking a median 61 months[1][2]
• Demand is booked, not forecast: $2.33 trillion of contracted future revenue sits behind $730B of guided 2026 hyperscaler capex; Microsoft, Alphabet, Amazon and Meta all state demand exceeds available capacity[2]
• Prices prove it: the first hyperscaler GPU price increase in 20 years (AWS H200 instance +15% in January 2026), Blackwell rentals +48% in two months, H100 lead times of a year, aging A100s appreciating[3][11][14]
• But the shortage is unevenly distributed — a "computing divide" where five hyperscalers control ~two-thirds of global AI capacity; startups pay 2–3× more and get throttled, academia is priced out of frontier training entirely[15][16][19]
• The skeptics have a real case too: xAI's Colossus ran at ~11% utilization (vs 35–45% for production training) and was rented out to rivals for ~$26B/year; Michael Burry argues $3T of off-balance-sheet commitments plus over-long depreciation schedules mask an eventual bust; Goldman Sachs expects supply-demand balance to moderate from 2027[13][20][21][22]
• Verdict: a genuine, quantified shortage through at least 2027–28, concentrated in memory and power; the risk is not that compute doesn't exist but that access to it has become the moat — and that the bill (est. $5.2–7.9 trillion by 2030) is being financed before demand is proven[4][12][23]

Part 1 — What "compute" actually is (and why it's physical)

Compute = the CPU/GPU capacity to train and run AI models. It depends on a deeply intertwined physical supply chain: accelerators, high-bandwidth memory, networking equipment, large-scale data centres, and enormous amounts of electricity. AI appears digital but is increasingly constrained by the physical world — chips, memory, electricity, land, cooling and industrial infrastructure[1].

The core numbers (Epoch AI, CC BY 4.0):

Demand grows roughly 7× faster than the slowest supply layer — energy. The effective supply growth is the minimum of all layers (Liebig's law), and the supply-demand gap is close to 300 percentage points per year[7].

Part 2 — The demand side: booked, not speculative

• Hyperscaler capex 2026: Microsoft $140.6B (+65%), Alphabet $195–205B (raised twice this year), Amazon ~$220B, Meta $130–145B. Behind it: $2.33T of contracted future revenue — booked orders, not forecasts[2]
• OpenAI has committed ~$750B of compute through 2030; Anthropic ~$65B, bought before the capacity exists[2]
• The frontier race is explicitly a compute race: an OpenAI investor memo (April 2026) brags it plans "30 gigawatts of compute by 2030" vs Anthropic's expected "7 to 8 gigawatts by end-2027"[9]
• Demand drivers compound: training compute for frontier models grows ~5×/year; the global AI chip stock doubles every ~7 months (3.4×/year); agentic AI runs inference 24/7 — agents use far more tokens than chatbot requests[1][8]
• a16z internal data: token consumption grew 100× from March to August 2026; fewer than 10 million heavy paying AI users exist today against 1.5 billion knowledge workers not yet adopted[12]

Part 3 — The four simultaneous bottlenecks

1. GPUs. On-demand capacity effectively sold out. H100/H200 lead times 36–52 weeks; Blackwell pricing +15–23%; even older-generation rental rates rising to 2024 peaks. Nvidia guided $108B revenue for a single quarter — framed by Jensen Huang as a supply number against ~100% demand[1][2][11]

2. Memory (the binding chip constraint). HBM takes ~3× the wafer capacity of standard DRAM per gigabyte, with no new capacity before 2027. AMD became the first company in history to write "industry-wide memory shortage" into a risk factor; Intel called it "one of the most severe supply constraints in its history". Spot DRAM prices ~8× since early 2025; SK hynix's CEO: "2027 will represent the worst supply shortage in the memory industry's history"[1][2][11]

3. Silicon production. TSMC's advanced-node capacity (N3, which underpins the AI accelerator ecosystem) is approaching full utilization through at least 2027; advanced fabs and the EUV machines to equip them take years to build[1]

4. Electricity (the structural constraint). Silicon is cyclical; power is structural. Data-centre power demand forecasts imply 200–300 GW globally by 2030; GE Vernova and Siemens Energy are nearly sold out of gas turbines through 2029; median grid-connection lead time is 61 months; the PJM capacity auction cleared at its ceiling and still fell 6,623 MW short[1][2]

The single most telling data point: in January 2026 AWS quietly raised the price of its 8×H200 instance from $34.61 to $39.80/hour — the first meaningful hyperscaler GPU price increase since EC2 launched in 2006. Twenty years of compute deflation ended with a silent line-item update[11].

Part 4 — What industry voices say

Part 5 — Who the shortage actually hurts

The shortage is not equally distributed — it is a computing divide:

• Concentration: five hyperscalers (Google, Microsoft, Meta, Amazon, Oracle) control roughly two-thirds of world AI compute; OpenAI alone uses an estimated 10–15% of operational global compute[15]
• Startups ("neolabs") pay 2–3× more for on-demand cloud GPUs, get throttled at peak times, and are now renegotiating expiring 2–3-year contracts into much worse terms. Evan Morikawa (Generalist) contacted 17 different providers to secure compute for robotics model training. AI-first startups spend 40–50% of revenue on compute, with gross margins of 25–60% vs 75–85% for traditional software[16][17][19]
• Academia is priced out of frontier training entirely — shared university clusters mean days-to-weeks of wait times on older hardware; research questions increasingly chosen to fit available compute[15]
• The $500B+ financing wave makes it worse, not better — when the biggest buyers borrow at attractive rates, the marginal price for everyone else rises. The compute divide "isn't closing — it's being financed into permanence"[15]
• Even NVIDIA's response — a revenue-sharing program trading startup equity for GPU access — was reportedly paused within weeks over antitrust concerns, an admission of how concentrated access has become[18]

Part 6 — The skeptical case: shortage or bubble?

A serious challenge to the shortage narrative exists:

1. Underutilization is real. xAI's Colossus 1 ran at ~11% Model FLOPs Utilization (vs 35–45% for production training) — "embarrassingly low" per its own internal memo. Rather than fix it, xAI rented the whole 220,000-GPU machine to rivals Anthropic (~$1.25B/month) and Google (~$920M/month) — roughly $26B/year of compute rented out by a frontier lab, which reads as evidence of both scarcity (at the margin) and slack (in the middle)[13]
2. Financial engineering, not demand. Michael Burry: nine tech giants carry ~$3T in off-balance-sheet commitments ($1.2T unstarted leases + $1.9T purchase commitments) vs $600B of actual capex; hyperscalers under-depreciate chips (5–6-year schedules on 2–3-year economic lives), potentially suppressing ~$176B of expenses 2026–28; and circular financing (chipmaker invests in customer, customer buys chips) can "artificially reinforce demand and revenue"[20]
3. The contradiction at the heart of the trade. Burry: "The hyperscalers are promising permanent demand growth and temporary spending over 3–4 years all in the same breath. Nvidia needs the permanent demand growth. Hyperscalers need the buildout to be largely over in 3–4 years... There is no third door."[21]
4. Supply is coming. Goldman Sachs expects supply-demand balance to tighten through 2026 then moderate from 2027 as new capacity comes online; packaging shortages narrow from ~20% to ~10% by end-2026 and stop binding during 2027; McKinsey's constrained-demand scenario needs "only" $3.7T rather than $5.2T[2][4][22]
5. Efficiency (Jevons paradox cuts both ways). The 2025 DeepSeek shock briefly crashed Nvidia $600B in a day on the theory that efficiency kills demand. The dominant counter-reading: efficiency increases total consumption (Nadella: "Jevons paradox strikes again!"), which is why capex went up after DeepSeek[10]

Part 7 — What the neutral forecasters actually project

• Epoch AI: training-compute demand grows ~5×/year; the largest 2030 training runs will draw 4–16 GW each; >100 GW of total AI capacity worldwide plausible by 2030 (approaching 5% of US electricity). Crucially, Epoch's "Can AI scaling continue?" study finds the bottlenecks — chips, data, power — will not hard-stop scaling by 2030: 10,000×-GPT-4-scale runs remain feasible if companies pay for them[6][8][23]
• IEA: data-centre electricity ~415 TWh in 2024 (1.5% of global) → ~945 TWh by 2030 (~3%); accelerated-server (AI) demand grows 30%/year; scenario range 700–1,700+ TWh by 2035[5]
• Goldman Sachs: 55 GW today → 84 GW by 2027 → ~122 GW by 2030, with the balance moderating from 2027[22]
• McKinsey: ~$6.7T cumulative capex needed by 2030 ($5.2T AI-related); demand could nearly triple by 2030 with ~70% from AI workloads[4]

Note the tension: the long-run forecasts say enough supply is being built; the 2026 spot market says there isn't enough right now. Both are true — the shortage is a transition-state phenomenon between 100-year-old power infrastructure and a demand curve growing 300%/year.

Part 8 — Verdict

There is a shortage, and it is verified, priced and contractual — not narrative. The evidence: allocation queues instead of spot sales, 20 years of deflation ending in a single AWS pricing update, rental prices up double digits, every major buyer stating demand exceeds capacity, and $2.33T of booked revenue behind the buildout[2][11][14].

But three qualifications matter as much as the verdict:

1. The binding shortage is memory and power, not GPUs per se — and power is the one constraint a purchase order cannot accelerate[1][2]
2. It is a shortage of access, not just capacity — idle and underutilized GPUs coexist with rationing at the margin (xAI's 11% utilization machine rented out at $26B/year); ownership, not merit, decides which ideas reach the market[13][17]
3. It has a shelf life — consensus paths to normalization run 2027–2029 (memory capacity lands, packaging unbinds, Goldman sees moderation from 2027), and the bear case (Burry's $3T footnote bill, depreciation games, circular financing) is a claim about what happens after the shortage, not whether it exists[2][20][21][22]

The honest summary: through 2026–2028 compute is what oil was in the 1970s — genuinely scarce, strategically hoarded, and the deciding factor in who can compete. Whether that ends in an infrastructure dividend or a $3 trillion write-down is the single biggest open question in technology.

Sources

1. Apollo — The Growing Compute Shortage — https://www.apollo.com/institutional/insights-news/insights/2026/06/growing-compute-shortage
2. sec-api.io — AI infrastructure supply and demand: what 42 issuers disclosed — https://sec-api.io/insights/ai-infrastructure-supply-and-demand-what-42-issuers-disclosed
3. Aethir — Inside the 2026 GPU Shortage — https://aethir.com/blog-posts/inside-the-2026-gpu-shortage-compute-demand-is-rapidly-growing
4. McKinsey — The cost of compute: a $7 trillion race to scale data centers — https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-cost-of-compute-a-7-trillion-dollar-race-to-scale-data-centers
5. IEA — Energy demand from AI — https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai
6. Epoch AI — How much power will frontier AI training demand in 2030? — https://epoch.ai/blog/power-demands-of-frontier-ai-training
7. Stanford MS&E 435 — AI Compute Demand (Epoch AI data, CC BY 4.0) — https://mse435.stanford.edu/ai-compute-demand.html
8. Epoch AI — trends dashboard (compute stock, training runs) — https://epoch.ai/
9. CNBC — OpenAI memo: 30 GW by 2030 vs Anthropic 7–8 GW — https://www.cnbc.com/2026/04/09/openai-slams-anthropic-in-memo-to-shareholders-as-rival-gains-momentum.html
10. Forbes (Jon Markman) — The Jevons Paradox: flawed consensus view on efficiency — https://www.forbes.com/sites/jonmarkman/2026/01/27/the-jevons-paradox-flawed-consensus-view-on-efficiency/
11. NeuralWired — NVIDIA GPU shortage 2026: who gets priced out — https://neuralwired.com/2026/05/30/nvidia-gpu-shortage-2026-ai-compute
12. BigGo Finance — Gavin Baker: AI compute shortage will last through 2028 — https://finance.biggo.com/news/5483e0ee7fc171b5
13. Hanwil Holdings — The AI capex boom revisited (xAI Colossus underutilization) — https://hanwilholdings.substack.com/p/the-ai-capex-boom-revisited-is-capital
14. Tech & Business News Weekly — Blackwell GPU rentals +48% — https://risenshine63.blogspot.com/2026/04/tech-ai-weekly-wrap-april-12-19-2026.html
15. Offworld News — The compute divide that money won't close — https://offworldnews.ai/half-a-trillion-more-the-compute-divide-that-money-wont-close/
16. VFF — Why the AI compute crunch is hitting neolabs especially hard — https://vff.ai/article/2026/08/13/why-the-ai-compute-crunch-is-hitting-neolabs-especially-hard
17. IEEE ComSoc — AI compute has a switchboard problem — https://techblog.comsoc.org/2026/09/01/ai-compute-has-a-switchboard-problem-orchestration-plane-needed/
18. Forward Future — The GPU access problem NVIDIA can't solve alone — https://forwardfuture.com/newsletter/originals/the-gpu-access-problem-nvidia-can-t-solve-alone
19. Digital Today — AI computing power shortage deepens, squeezing AI startups — https://www.digitaltoday.co.kr/en/view/51283/ai-computing-power-shortage-deepens-pressure-on-ai-startups
20. Yahoo Finance — Michael Burry: Big Tech's hidden $3T tab — https://finance.yahoo.com/technology/ai/articles/michael-burry-says-could-heard-055211614.html
21. Yahoo Finance — Burry: there is no third door — https://sg.finance.yahoo.com/news/big-short-investor-michael-burry-164825420.html
22. Goldman Sachs Research — AI to drive 165% increase in data center power demand by 2030 — https://www.goldmansachs.com/insights/articles/ai-to-drive-165-increase-in-data-center-power-demand-by-2030
23. Epoch AI — Can AI scaling continue through 2030? — https://epoch.ai/publications/can-ai-scaling-continue-through-2030
24. The Guardian (via AOL) — Microsoft's AI plans and the chip-count discrepancy — https://www.aol.co.uk/articles/microsoft-ai-plans-being-held-040046000.html