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

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

Alex Kantrowitz speaks with Dylan Patel and Jordan Nanos of SemiAnalysis about the AI infrastructure boom, whether today’s massive capex can be justified, and what could slow it down. The conversation centers on the economics of AI data centers, the role of NVIDIA and the neoclouds, the profitability of AI labs like Anthropic and OpenAI, and the operational reality behind the hype — including security, reliability, and execution risk.

The AI infrastructure buildout is bigger than most people realize

• Dylan Patel argues the current AI capex cycle is already around 5%–6% of U.S. GDP, higher than the commonly cited figures.
• He says the true scale is understated because:
  • AI spending includes not just chips and data centers, but the broader supply chain.
  • U.S.-based infrastructure is often serving global inference demand.
• His core point: this is not just another hardware cycle — it may be the largest infrastructure buildout in history.

Will spending growth keep going up forever?

• Jordan says the buildout will likely hit limits toward the end of next year.
• The big constraint is not just demand, but:
  • difficulty building additional gigawatts of capacity,
  • securing enough capital,
  • and the fact that growth rates inevitably slow as the base gets larger.
• Dylan agrees that spending growth must decelerate eventually, but he emphasizes that demand is still outrunning supply.

The “how do we justify it?” debate

The hosts discuss projections from Goldman Sachs, Bain, and others arguing AI will need trillions in annual revenue to pay back the infrastructure.

• Dylan’s response: those estimates are likely too low.
• He argues revenue trails capex naturally because:
  • GPU/data center investments are depreciated over years,
  • and current infrastructure spending is laying the groundwork for future revenue.
• His view is that the market is underestimating how large AI revenue could become by 2027–2031.

AI labs may already be approaching break-even

• Dylan says companies like Anthropic are already close to or at profitability on compute vs. revenue.
• He also claims:
  • Anthropic could reach hundreds of billions in revenue by late 2026,
  • OpenAI could get there shortly after.
• His broader thesis is that AI economics are not like old software economics:
  • margins can still be strong,
  • customer acquisition costs can be low,
  • and the customer base is sticky because switching models is difficult.

Where the money comes from

A major question is: if AI services become a huge share of GDP, who pays?

• Dylan’s answer:
  • companies will reallocate more of their R&D and COGS spending toward AI,
  • AI will help firms do more with fewer people,
  • and the rest of the economy may grow fast enough that labor income does not collapse immediately.
• He frames AI as accelerating a long-term trend:
  • capital share rises
  • labor share falls
• He also points to industries where AI could create major new revenue:
  • chip design
  • robotics
  • trading
  • pharma and drug discovery
  • self-driving systems

Risks: project cancellation, demand slowdown, and solvency

• The big near-term risk is not that Meta or others can instantly cancel all contracts.
• Jordan clarifies that:
  • many hyperscaler/neocloud deals are multi-year term contracts and hard to walk away from,
  • but companies can slow or pause self-build projects if needed.
• Dylan says the real question is whether demand slows enough to create a supply-demand mismatch.
• For now, the answer is no: demand still appears to exceed supply.

Anthropic IPO, valuation, and “if it’s worth $20T, society may be in trouble”

• Dylan’s provocative tweet about an Anthropic IPO at a multi-trillion valuation leads to a broader discussion.
• His point:
  • if Anthropic becomes a $20T company, AI is likely so economically powerful that society may be radically reshaped.
• That could mean:
  • huge productivity gains,
  • or severe concentration of wealth and social unrest if gains accrue to a tiny group.
• Jordan’s take is more optimistic:
  • if Anthropic reaches that scale, it likely isn’t just selling tokens,
  • it is probably deploying AI across major industries like medicine and biotech.

Why Anthropic can look like a great business

Dylan argues Anthropic may actually be better than traditional SaaS in some ways:

• Gross margins on inference are around 75%+
• Customer acquisition costs are low
• Revenue is growing fast
• The business appears sticky

His key claim: AI labs can have strong economics because they sit at the center of the compute stack and can capture value across many industries, not just via API tokens.

ClusterMax, neoclouds, and the quality problem

A large portion of the episode is devoted to SemiAnalysis’s ClusterMax 3.0, which ranks GPU cloud providers by operational quality.

What ClusterMax measures

• Reliability
• Security
• Storage/network quality
• Lifecycle management
• Slurm/Kubernetes operations
• Overall usability for customers renting GPU capacity

Top-tier providers

• CoreWeave
• Nebius
• Oracle
• Google Cloud

Main takeaway

• The top providers are strong, but many others are failing on basics.
• Jordan says it is disappointing how many clusters still have weak security and reliability despite the amount of money flowing into the space.
• The market is still a seller’s market, so providers focus more on raising capital and building fast than on perfect service quality.

NVIDIA’s role as a backstop

Jordan explains that NVIDIA is helping support the ecosystem through:

• revenue floors,
• landlord guarantees,
• lease structures,
• and other off-balance-sheet backstops.

He cites a figure of roughly $588 billion in NVIDIA-supported backstops.

Why NVIDIA does this

• To diversify its customer base
• To reduce dependence on hyperscalers
• To ensure the market has enough financing to keep building

The hosts frame this as pragmatic rather than sinister: NVIDIA wants a healthy, liquid ecosystem for GPU deployment.

Security risks are real

The episode spends serious time on cybersecurity in neoclouds.

• Ilya Sutskever’s warning is cited: rogue agents could try to hijack neoclouds to run copies of themselves.
• Dylan says many neoclouds are poorly secured.
• He mentions examples where outsiders could potentially view other customers’ storage and job data.
• The takeaway: AI infrastructure security is still immature, even as the industry scales.

Oracle, Google, and the changing cloud hierarchy

Oracle

• Jordan says Oracle ranks highly in ClusterMax on service quality.
• But that does not mean the stock is safe.
• Oracle has faced:
  • debt concerns,
  • project delays,
  • regulatory / infrastructure issues like pipeline constraints in New Mexico.
• So: strong cluster operator, but still exposed to execution and financing risk.

Google

• Google is increasingly allocating compute away from DeepMind and toward commercial monetization.
• Jordan notes DeepMind now appears to have less compute than Anthropic or OpenAI by year-end.
• The implication:
  • Google is becoming more like a full-stack AI business,
  • but its frontier research position may be weakening unless it sharpens focus.

Will frontier labs keep models internal?

The episode ends with a discussion of whether labs will keep better models in-house longer.

• Dylan says Anthropic and OpenAI are already keeping some newer models internal.
• Reason:
  • releasing models publicly creates distillation and competitive risk.
• Jordan’s view:
  • labs may continue releasing only the minimum needed to stay ahead of Chinese open-source competitors.
• The tension:
  • release more and monetize via API,
  • or keep models internal and use them to build higher-value products.

Main takeaways

• The AI infrastructure boom is larger than most people think, and may already be at several percent of GDP.
• The buildout is real, but growth rates must decelerate eventually.
• AI labs may be closer to profitability than skeptics believe.
• The biggest risks are not just financial — they include security, execution, and concentration of power.
• SemiAnalysis’s ClusterMax suggests that only a few providers are truly ready for the frontier.
• NVIDIA, Oracle, Google, OpenAI, and Anthropic are all playing different roles in a rapidly evolving ecosystem that still has room to expand dramatically.