AI’s $3 trillion problem is really a demand problem

Three years ago, Sequoia’s David Cahn put a blunt number on the AI buildout: if Nvidia was already pulling in about $50 billion a year from GPUs, the industry would need roughly $200 billion in revenue just to justify the infrastructure wave underway.

That estimate aged fast. Cahn now pegs 2026 AI infrastructure spending at about $1.5 trillion. On his math, the industry needs around $3 trillion in revenue to pay for the chips, data centers, power, networking, and everything else in the stack.

That’s a huge number, but the bigger point is simpler. The constraint isn’t whether the hardware exists or the models can be trained. It’s whether software and services can consume enough compute, often enough, at prices customers will actually pay.

The math behind the panic

The AI money machine looks straightforward from far away. Hyperscalers buy accelerators, build giant clusters, rent them through cloud APIs, and hope usage climbs fast enough to justify the capital outlay. If enough companies wire AI into core workflows, the spend gets absorbed. If not, the capex turns into a very expensive bet on demand that never quite shows up.

Cahn’s updated estimate matters because the cost mix has changed so quickly. The early AI race was framed around training. That was the part everyone talked about. Now the bill is dominated by everything around it: inference fleets, high-bandwidth memory, cooling, power delivery, specialized networking, and the data centers themselves.

That’s where the economics get ugly. Newer chips are faster, but they’re also pricier. Memory is a real constraint. Inference-specific silicon is eating into the idea that one general-purpose GPU can do everything. Each bottleneck pushes the revenue needed to make the buildout pay off even higher.

A lot of AI product teams still miss this. A model can be impressive and still be uneconomic. A system can cut latency and still make the bill worse if usage spikes. Efficiency improvements are good for customers. They can be bad for the vendors whose business depends on high token throughput.

Cheaper tokens help users, not the bill

Torsten Slok, chief economist at Apollo, has been pointing at the same gap from a different angle. The hyperscalers, he notes, are effectively assuming free cash flow will accelerate sharply by 2028, which is when they expect the payback from all this infrastructure spending to show up.

That assumption gets shaky if model usage keeps getting cheaper for everyone but the sellers of compute.

Right now, a few forces are pushing token prices down:

• More organizations are using open-weight models, including a lot of Chinese models that are cheaper than frontier lab offerings.
• Frontier models are getting more efficient. OpenAI’s latest model is reportedly 54% more token efficient on coding tasks, according to Sam Altman.
• Customers are getting better at routing. They use large models only when they need them, then fall back to smaller, cheaper ones for routine work.

For developers, that’s good news. The cost of building agents, copilots, search systems, and internal automation keeps falling. You can ship more with less budget. You can also run more evals, more retries, more tool calls. That matters.

For the infrastructure vendors, it’s a tougher story. Lower token prices don’t automatically mean higher total usage. The bet depends on a familiar cloud pattern: cheaper units lead to so much more consumption that total revenue still rises. That may happen. It may also stall if AI settles into narrower, more predictable enterprise use cases.

That’s the real risk. Not that AI stops working. It already works. The risk is that it works well enough to get embedded, but not broadly enough to drive the usage levels needed to pay the bill.

Why open models matter

Open-weight models have been changing buyer behavior for a while, but the effect is getting easier to see. If a company can run a capable model from a vendor with lower pricing, or host it itself, frontier lab pricing starts to look like a convenience tax.

That has direct technical implications.

A team choosing between a proprietary API and an open-weight model isn’t just comparing benchmark scores. It’s comparing:

• latency and throughput under load
• fine-tuning and control
• data residency and compliance
• failure modes and observability
• total cost at scale, including cache hits, retries, and orchestration overhead

The cheapest model on paper often loses once you add the work around it. But expensive models lose too when teams can get 80% of the quality for 20% of the cost. That trade-off is getting sharper as open models improve and inference stacks get better at serving them.

There’s also a strategic wrinkle. If enough enterprise buyers move to cheaper models, frontier labs lose pricing power. If frontier labs respond by cutting prices, hyperscaler revenue math gets harder. Either way, the market starts pressing on the same assumption: that a small number of model providers can keep extracting huge margins from AI usage.

The hyperscaler bet is narrower than it looks

Google, Meta, Microsoft, and Amazon can spend at a scale most companies can barely model in a spreadsheet. That’s exactly why they matter here. So much AI capital is concentrated in so few names that their spending and their earnings guidance are now linked.

If those companies miss their expected payback window, the problem doesn’t stay inside “AI stocks.” Apollo’s Slok is right to flag the macro risk. When a large share of market cap depends on AI infrastructure monetizing on schedule, a slowdown becomes a broader financial event.

The current capex cycle has a few fragile assumptions baked into it:

• demand keeps rising fast enough to absorb new capacity
• inference economics improve before pricing gets crushed
• enterprise adoption scales beyond pilot projects and novelty use cases
• model quality improvements translate into budget expansion, not just cost savings

The catch is that all of those can be true in isolation and still fail together. A company might adopt AI more broadly while spending less per task. A model might get better while using fewer tokens. A cloud provider might fill more racks while earning less per unit. That’s good engineering. It’s bad revenue growth.

What developers should watch

If you’re building AI systems, the headline number matters less than the signals underneath it.

Watch token economics the way you’d watch query costs in a database product. Model choice is becoming a routing problem, not a philosophical one. The best architecture increasingly looks like:

• small models for classification, extraction, and routing
• larger models for hard reasoning and code generation
• caching and reuse wherever output is deterministic enough
• evals tied to business cost, not just benchmark scores

That setup can keep product margins sane. It also means teams need better instrumentation. You can’t manage what you don’t measure, and AI systems are good at hiding cost in retries, long context windows, tool calls, and badly designed agent loops.

This is where a lot of teams still get burned. They optimize for model quality, then find that 10% better output comes with 4x the cost. Or they chase the cheapest model, then spend the savings on human review because reliability falls apart. The right answer is usually a mixed stack with hard guardrails and routing logic, not one model everywhere.

There’s a security angle too. As more organizations move to open models or self-hosted inference, the attack surface changes. You inherit patching, model provenance issues, supply-chain risks, and the reality of running CUDA, drivers, kernels, and orchestration software at scale. Cheap tokens can get expensive fast if ops discipline is weak.

The bill is real

A lot of AI commentary still treats infrastructure spend like a sign of confidence. It is that. It’s also a liability that has to clear a very high revenue hurdle.

Cahn’s $3 trillion number is useful because it forces the question investors and builders keep trying to dodge: who pays for all this, and at what margin? Not whether AI matters. Not whether the tech is impressive. Those answers are obvious. The question is whether the industry can turn an enormous amount of compute into an even larger amount of cash.

Right now, cheaper tokens are great for users and a headache for the people financing the grid underneath them. That tension is only getting sharper.