Can AI Keep Growing Exponentially? Let’s Ask SemiAnalysis — With Dylan Patel and Jordan Nanos

Summary

• SemiAnalysis argues that viral charts materially understate the scale of the AI buildout. Dylan Patel estimates next-year US capital spending near $2 trillion across data centers, chips, and the supporting supply chain—roughly 5–6% of a $30-trillion-plus economy, not 3.6%. America is also building training and inference capacity for overseas users, making this “much greater than in any other cycle of physical infrastructure investment.”
• Triple-digit capex growth must slow, but SemiAnalysis does not yet see the demand signal that would mark a bust. Hyperscaler capex was growing 116% year over year in Q3 2026, while engineering capacity, cooling, chips, capital, and the ability to organize another gigawatt are becoming constraints. Demand still “significantly exceeds supply,” and nobody can tell whether the excess is 10%, 50%, or effectively unbounded until enough capacity comes online.
• The payback math looks less absurd when infrastructure life and revenue lag are included. Dylan models GPUs over six years and data centers over 15: $1 trillion invested today might require about $1.5 trillion of cumulative revenue, or $250 billion annually over six years—not immediate matching revenue. He thinks only $3 trillion of AI revenue in 2031 would be a problem, while suggesting OpenAI and Anthropic together might reach $600–700 million by December of next year and perhaps $1.5 trillion by the end of 2027.
• AI spending is expected to draw on labor, R&D, and cost of goods sold while shifting value toward compute-based capital. Dylan’s formulation is that tokens are “essentially just capital” embodied in chips and data centers; companies can keep headcount stable while producing and selling much more. His chip-design example is the template: US engineering headcount stayed broadly flat for two decades while tools helped create far more valuable chips across phones, sensors, robots, and AI systems.
• Anthropic could have software-like—or better—economics, but its upside case implies profound social disruption. Dylan says Anthropic already earns roughly 75% gross margins on output, has almost no customer-acquisition cost, and may be a “better business than software” if it sustains model leadership. Yet a $20 trillion valuation at perhaps 5× sales implies several trillion dollars of revenue and an AI economy so concentrated that, in his deliberately pessimistic framing, “society as a whole could be in trouble.”
• Neo-cloud financing is real financial plumbing, not sufficient evidence that NVIDIA fabricated demand. Jordan Nanos says SemiAnalysis tracks $588 billion of NVIDIA-backed off-balance-sheet guarantees, about 2.5% of US M2, which can make projects financeable by giving lenders an investment-grade fallback buyer. NVIDIA often sells the capacity to customers willing to pay more, while the broader constraint spans construction, memory, optics, factories, and industrial equipment—not merely GPU purchases.
• The near-term fault line is operational quality and security, while frontier labs increasingly reserve their strongest models for themselves. SemiAnalysis found only a small recommended set among 323 tracked providers and described cases where customers could see other tenants’ data and government workloads; “most of these clouds are terrible when it comes to security.” Meanwhile Anthropic and OpenAI have delayed or limited models including Mythos, Fable, Glasswing, Astra, and Bell, suggesting public releases may become “minimally viable models” that preserve an edge while higher-value capabilities stay internal.

Deep dive

1. The AI buildout is larger than the headline comparisons

• Dylan’s reaction to the viral infrastructure chart was that “people’s numbers are undercounted.” He puts next-year US spending near $2 trillion once data centers, chips, and the surrounding supply chain are included—about 5–6% of GDP rather than the chart’s 3.6%.

• Railways and fiber were largely built for the countries recording the investment. US AI infrastructure trains models domestically and serves inference abroad: Dylan cites European Anthropic and OpenAI usage, plus Asian demand, as reasons America’s build carries economic capacity beyond its borders.

• Alex’s benchmark was 116% year-over-year hyperscaler capex growth in Q3 2026, with consensus falling below 100% in Q4 and toward 70% in the first quarter of next year. Jordan says limits should arrive toward the end of next year and beyond because EPC capacity, industrial equipment, cooling, chips, and financing cannot keep compounding at triple-digit rates indefinitely.

2. Slower capex growth would not itself mean the boom is over

• The arithmetic is already becoming harder: Dylan describes revenue rising from below $10 billion to above $100 billion in one year, but says it will not become $1 trillion by the end of next year. NVIDIA can still double while visibly slowing from earlier multiples.

• SemiAnalysis’s distinction is crucial: infrastructure spending growth must eventually slow, while revenue and profit growth must accelerate to justify the installed base. Current AI revenue reflects prior years’ construction, so comparing today’s sales directly with today’s capex mismatches two different periods.

• Jordan’s demand test is concrete: do not forecast cancellations until usage slows. Every model presently shows demand “significantly exceeds supply,” but constrained GPU and data-center availability prevents anyone from knowing whether unmet demand is 10%, 50%, or far larger.

• The host’s neo-cloud concern prompted an important correction: major customers generally sign five- or six-year capacity contracts at fixed hourly rates, not cancellable monthly commitments. Meta can leave its own campuses empty or delay buying chips for them, but it cannot simply abandon signed third-party capacity without consequence.

3. The payback case depends on asset life and much broader monetization

• Dylan uses six years for GPU depreciation and roughly 15 years for data centers. On his simplified math, $1 trillion invested today might need $1.5 trillion over six years—around $250 billion annually—while the revenue curve could start low and rise sharply later.

• That is why he dismisses the claim that Meta committed “corporate suicide.” Meta has contractual obligations, but it can stop authorizing additional campuses; its existing cash flow can cover what it has signed, leaving the real question as return on investment rather than solvency.

• Dylan says only $3 trillion of AI revenue by 2031 would indeed create “a big problem.” He says OpenAI and Anthropic together could reach $600–700 million by December of next year and perhaps $1.5 trillion by the end of 2027. He says Anthropic is already profitable on revenue minus compute costs for training and inference, while OpenAI could follow within two or three quarters.

• Dylan argues that token subscriptions are too narrow a revenue frame. Frontier intelligence could enter robotics, autonomous vehicles, drug development, healthcare, and life sciences; a lab using AI to produce a cancer treatment could capture economics closer to a breakthrough drug than a cloud API.

4. AI shifts value from labor toward capital without necessarily shrinking payrolls

• Asked where hundreds of billions of customer spending would come from, Dylan points to expanding R&D budgets, cost of goods sold, and the labor force. Tokens are “essentially just capital”—chips and data centers performing work—so AI accelerates the decades-long decline in labor’s share of GDP.

• His explicitly hedged hope is that rapid overall growth keeps labor income stable or even rising in dollars despite its falling percentage share. Companies might employ the same number of people while building more products, charging more, improving margins, and directing a larger portion of incremental value to capital.

• Chip design is Dylan’s model for the transition: US R&D engineering headcount has barely changed for roughly 20 years, yet the sector’s value exploded as better tools amplified each developer. AI could repeat that pattern across chips for phones, AirPods, robots, GPS, Bluetooth, gas sensors, and AI itself.

• Alex pressed Dylan on his earlier suggestion that intellectual work was ending. Dylan separated the productive period through roughly 2028—when unusually active people gain enormous leverage—from the “madness of 2035,” when systems better than humans at everything might make intellectual labor genuinely obsolete.

5. Anthropic’s upside is both commercially attractive and socially destabilizing

• Dylan called his $2 trillion Anthropic IPO post “a complete trash post,” then defended its underlying paradox. Anthropic might become the first $10 trillion company, but $20 trillion at mature multiples near 5× sales could require $2–4 trillion of revenue and capabilities powerful enough to transform the global economy.

• His pessimistic branch is extreme concentration: employment disruption and gains accruing to Anthropic shareholders or supply-chain owners could produce political unrest and “tear apart the fabric of society.” Jordan is less pessimistic, imagining value created through new medicines and invoking GLP-1 as an example of research generating both broad benefits and exceptional profits.

• The operating case is already strong. Dylan says Anthropic earns around 75% gross margins on output, especially advanced models, with customer-acquisition cost “almost zero”; R&D is rising more slowly than revenue, while existing customers appear loyal and tend to expand spending.

• Claims of voluntarily “slowing down the frontier” strike Dylan as signaling more than strategy. Dario cannot trust Sam to stop, Sam cannot trust Dario to stop, and neither can ignore China; serious security incidents may cause temporary restraint, but “they have very little opportunity to slow down.”

6. NVIDIA’s guarantees finance a constrained market rather than manufacture it

• Alex described the report’s mechanism as including minimum-revenue, landlord, and lease guarantees. Jordan says SemiAnalysis tracks $588 billion of NVIDIA-backed off-balance-sheet guarantees, equivalent to about 2.5% of US M2. The mechanism gives lenders a credible fallback buyer, allowing neo-clouds to secure financing without relying exclusively on hyperscalers.

• In practice, NVIDIA may agree to take capacity but never need to: another customer arrives willing to pay more, so NVIDIA sells the GPUs to that party. Jordan therefore sees the arrangement as pragmatic diversification, giving NVIDIA more customers and labs more alternatives to three or four hyperscalers.

• NVIDIA also has legitimate internal demand through model research, software testing, and continuous integration, although Jordan says this consumes only a small share of its total business. More importantly, projects need financing for construction, memory, optics, factories, and equipment throughout the supply chain—not merely chips.

• The hard costs are similarly misunderstood: concrete, steel, and walls matter operationally but represent a minority of a new cluster’s bill. HBM, optics, and GPUs dominate, so an owner can complete a campus yet postpone its largest expenditure by leaving it unpopulated.

7. Cloud quality, security, and private models are becoming competitive moats

• ClusterMAX 3.0 evaluates the renter’s experience—reliability, security, storage, networking, lifecycle tooling, Slurm, and Kubernetes—not stock attractiveness. CoreWeave and Nebius rank highest, while Oracle can provide excellent managed clusters even as debt, fixed-margin contracts, and execution failures hurt its investment case.

• Oracle’s New Mexico site illustrates the split: missing pipelines led to delays and eventually force majeure, while Wisconsin also faced slippage. Nebius, with the shortest average contract duration among leading providers, could raise prices quickly; Oracle and CoreWeave were more constrained by long-term agreements.

• Of 323 tracked providers, only a few satisfy SemiAnalysis’s recommendation threshold. Security is the ugliest failure: Dylan says GLM 5.3 does a good job of hacking neo-clouds and that Chinese open-source models can hack dozens of them. He also recounts accidentally seeing other tenants’ data, Slurm activity, and GPU workloads belonging to agencies from a non-US top-10-GDP country.

• Google faces a different allocation problem. Jordan says DeepMind currently has the least computing power among DeepMind, Anthropic, and OpenAI, even as Google’s total compute rises. Dylan warns that next year OpenAI and Anthropic may add 5–10 gigawatts while Google adds only a few. “Nobody needs the third, fourth, or fifth best coding model,” he says of Gemini 4 Argon.

• Labs are already withholding frontier capability. Anthropic had Mythos ready in February; the public later got a limited Fable, while Glasswing was limited to some subsectors. OpenAI held Astra for months and displayed Bell’s math performance without releasing it. Dylan expects “minimally viable models” publicly—enough to stay ahead of Chinese and open-source systems—while months of progress remain internal.