SemiAnalysis Founder: AI Inference Could Outgrow Oil — What That Means for Crypto in the Next Compute Crunch
SemiAnalysis Founder: AI Inference Could Outgrow Oil — What That Means for Crypto in the Next Compute Crunch
AI is rapidly shifting from a “model training era” to an “inference era,” where the real economic gravity sits in running models billions (and soon trillions) of times per day. In a recent episode of Sequoia Capital’s podcast series, Training Data, SemiAnalysis founder Dylan Patel argued that AI inference could become one of the world’s largest markets — potentially bigger than oil — and large enough to represent multiple percentage points of global GDP. You can find the episode on the Training Data podcast feed.
For the blockchain industry, this isn’t just a macro tech story. It directly impacts decentralized compute, DePIN token economics, MEV and market structure, ZK verification workloads, and even how we think about wallet security in an agent-driven internet.
Below is a crypto-native reading of Patel’s thesis — and how builders and users can prepare for a world where compute behaves like a scarce commodity.
1) Why “Inference Is the New Oil” Matters to Web3
Patel’s core claim is not simply “AI is big.” It’s that inference demand expands faster than compute supply because each model iteration unlocks more tasks that are worth paying for. Even if chips get faster, and even if inference gets cheaper, demand can still accelerate.
This dynamic rhymes with something crypto users already understand:
- Lower costs don’t necessarily reduce total demand
Ethereum rollups and data-availability upgrades reduce per-transaction costs, but lower fees often invite more activity. Ethereum’s “rollup-centric roadmap” and data scaling path (including Proto-Danksharding / EIP-4844) is a canonical example of how cheaper “unit costs” can still lead to higher aggregate usage over time. See Ethereum’s overview of Danksharding and the scaling roadmap and the specification for EIP-4844.
Translate that to AI: if inference cost drops dramatically, usage can explode — and the system can remain constrained at the infrastructure layer.
2) A Long Compute Shortage Is a Centralization Risk (and a Crypto Opportunity)
In crypto, decentralization often fails for an unglamorous reason: bottlenecks.
- If blockspace is scarce, a few actors dominate MEV.
- If validator hardware becomes specialized, participation concentrates.
- If RPC access is expensive, developers route through a handful of providers.
Now add Patel’s view: compute scarcity may be persistent, not temporary. In that world:
Centralization risk: AI becomes a gatekeeper layer
If inference becomes “the new critical infrastructure,” then access to cheap, reliable inference can become a chokepoint for:
- trading and risk systems,
- wallets and user agents,
- fraud detection,
- compliance tooling,
- developer copilots that ship code.
Crypto opportunity: open markets for compute
Crypto is unusually good at coordinating markets where:
- suppliers are distributed,
- demand is global,
- payments are programmatic,
- and settlement is trust-minimized.
That’s why the rise of DePIN (Decentralized Physical Infrastructure Networks) matters. For a high-level framework, see a16z crypto’s explanation of what DePIN is and why it matters and Messari’s overview in its State of DePIN report (PDF).
The point isn’t that every DePIN network “wins.” The point is that blockchain offers a coordination substrate for provisioning real-world infrastructure (compute, bandwidth, storage, sensors) when incumbents can’t scale fast enough or can’t distribute access fairly.
3) Hardware–Software Co-Design: The Hidden Moat That Crypto Builders Should Notice
One of the most practical takeaways from Patel’s comments is that efficiency gains have increasingly come from cross-layer optimization (model architecture, kernels, runtimes, memory strategies), not only from new silicon. The Training Data episode description highlights this emphasis on co-design and ecosystem-level optimization. See the episode listing on Apple Podcasts.
For crypto, this changes how to think about “decentralized AI compute”:
- It’s not enough to aggregate GPUs.
- The winning systems will also ship:
- optimized inference stacks,
- scheduling that matches model shapes to hardware,
- and reproducible benchmarks that discourage “marketing performance.”
Patel’s team built a continuously updated inference benchmark called InferenceX, designed to measure real serving performance and cost. You can review how the project is positioned on the InferenceX site.
The crypto angle: verification becomes more valuable when execution is heterogeneous
As inference runs across diverse providers (hyperscalers, neo-clouds, DePIN networks, edge devices), users will want stronger guarantees:
- Did the provider run the model they claimed?
- Did they use the settings they claimed?
- Did they tamper with outputs?
- Can an on-chain system safely depend on off-chain inference?
This is where “verifiable compute” themes (including ZK proof systems and trusted execution approaches) become strategically important — even if most inference never happens on-chain.
4) Energy, Geopolitics, and the “Space Datacenter” Thought Experiment
Patel also points to energy cost and datacenter build capacity as core constraints — not just chips. This framing is consistent with the broader reality that AI is now an energy planning problem as much as a software problem.
For a credible macro view, the International Energy Agency’s report Energy and AI provides context on how AI and datacenters shape electricity demand. Start with the IEA’s executive summary and its section on energy demand from AI.
In Patel’s more forward-looking scenarios, he projects that:
- by 2030, the combined compute demand of just OpenAI and Anthropic could exceed 100 GW, and
- by 2040, a significant share of incremental compute could plausibly migrate off-planet if economics flip (space deployment becomes cheaper than terrestrial buildout).
A summary of these particular projections is widely circulated in industry recaps, for example this digest-style writeup.
Separately, SemiAnalysis has also explored the economics of orbital compute in its own work, including its analysis on space datacenters.
Why crypto should care (even if space compute sounds far off)
Whether or not you believe in orbital datacenters, the underlying point is immediate:
- Compute is becoming location-constrained by energy and permitting.
- Location constraints create jurisdictional constraints.
- Jurisdictional constraints create censorship and access constraints.
And censorship resistance is exactly what crypto was built for.
If AI inference becomes a foundational layer for finance and identity, then open, censorship-resistant settlement (blockchains) becomes more valuable — not less — because the world above it becomes more permissioned.
5) AI Agents + Wallets: “Who Holds the Keys?” Becomes the Real UX Question
As AI agents become more autonomous, crypto’s biggest mainstream UX unlock may also become its biggest risk: bots that can pay.
That trend is already visible in how investors and researchers discuss “agent wallets” and the legal/operational implications. See CoinDesk’s coverage referencing Electric Capital on wallets for AI agents and the emerging legal frontier.
If inference really does become “oil-scale,” then agents won’t just answer questions — they will:
- rebalance portfolios,
- post margin,
- stream payments,
- sign governance votes,
- rotate positions across chains,
- and interact with DeFi protocols continuously.
In that environment, self-custody is not optional. It’s the boundary between:
- an agent that can “suggest,” and
- an agent that can irreversibly move value.
Practical security pattern: keep high-value signing offline
For long-term holders, treasuries, and anyone delegating partial autonomy to software, the safest pattern is to separate:
- online intelligence (agents, inference, automation) from
- offline authority (final signing).
That’s where a hardware wallet can act as the human-controlled checkpoint. If you use OneKey, the core value proposition aligns naturally with this shift: private keys remain offline, while you can still participate across major chains and dApps without turning your seed phrase into an API key.
6) What Crypto Builders Should Do Next (A Short Checklist)
If Patel’s “inference-dominant economy” arrives faster than expected, crypto teams should prepare for a world where compute is the limiting reagent:
-
Assume compute becomes a pricing input
Don’t design token models that only work when inference is “effectively free.” -
Treat benchmarking as governance, not marketing
Transparent measurement (e.g., continuous inference benchmarks) will shape market trust more than whitepapers. -
Design for verifiability
If your protocol depends on off-chain AI, plan for adversarial providers and model drift. -
Build with energy realism
Energy and datacenter constraints will influence where infrastructure can exist — and therefore who can access it. -
Secure the signing path
As agents become more capable, users will need stronger signing controls, rate limits, multisig policies, and offline approvals.
Closing Thought: Crypto’s Job Is to Keep Markets Open — Even When Compute Isn’t
Patel’s most important message for the crypto industry may be this: even spectacular efficiency gains don’t guarantee abundance. If AI inference becomes a market measured in gigawatts and GDP points, then compute access will shape who gets to build, who gets to compete, and who gets priced out.
Blockchains can’t manufacture GPUs — but they can keep the economic layer open: permissionless markets, transparent settlement, and user-controlled keys. And in an agent-driven world, pairing that openness with strong self-custody practices (including an offline signer like OneKey) is how users keep sovereignty when the rest of the stack becomes scarce.
