Is the AI industry profitable? Yes, just not where you’re looking

The question “Is the AI industry profitable?” has two correct answers, and they point in opposite directions. At the chip-design and leading-edge-fabrication layers, AI is already one of the most profitable industries in commercial history. At the layers the market calls “AI”, frontier model labs, GPU-rental builders, and most applications built on someone else’s model, it is among the most loss-making activities ever financed.

Both statements are true. The investment question sits in the distance between them.

The reason is mechanical. Across the AI stack, the cost of intelligence is falling rapidly. But a falling cost only becomes profit somewhere. Whether that decline lands as margin, as a lower price to the customer, or as a transfer to a supplier depends on one question: who owns the bottleneck between the falling cost and the price the customer will pay?

Where a firm owns that bottleneck, it keeps the cost decline as margin. Where it owns none, competition forces the decline through. Walk the AI value chain from chips to applications, and the pattern is already visible. Profit sits where a pass-through is blocked. It evaporates where competition lets pass-through run free.

Start with the two companies that keep the money. NVIDIA reported fiscal-2026 revenue of US$215.9 billion, GAAP operating income of US$130.4 billion, and GAAP net income of US$120.1 billion, a net margin of nearly 56 per cent. TSMC earned 2025 net income of US$55.2 billion on revenue of US$122.4 billion, a 45.1 per cent net margin. Together, Nvidia and TSMC booked roughly US$175 billion of net profit in their latest fiscal years.

This is not a forecast. It is where AI profitability already exists.

Both companies sit behind gates that the rest of the stack must pass through. NVIDIA’s moat rests on CUDA, networking, scale, and the difficulty of coordinating around an alternative. TSMC’s moat is harder still: leading-edge fabrication is gated by physics, capital, yield learning, and process knowledge that takes years to reproduce. These are not normal suppliers. They are toll collectors.

The cloud layer is more ambiguous. AWS, Microsoft Azure, and Google Cloud are large, profitable businesses. AWS earned a 37.7 per cent operating margin in the first quarter of 2026, and Microsoft’s Intelligent Cloud has run margins in the low 40s. But hyperscaler free cash flow is being consumed by the AI build-out, and the cloud owners are trying to escape Nvidia’s toll through custom silicon. Amazon’s Trainium, Google’s TPUs, and Microsoft’s Maia are attempts to become bottleneck owners rather than resellers of someone else’s bottleneck.

Also Read: The AI economy is moving faster than our institutions

Where a cloud owner runs its own silicon, it can keep more margin. Where it buys Nvidia capacity, finances data centres, and rents compute to model labs, its economics compress. The cloud business is profitable, but AI infrastructure may not be unless demand arrives fast enough and custom silicon works well enough.

The neoclouds show what happens when revenue grows without a bottleneck. CoreWeave more than doubled first-quarter 2026 revenue to US$2.08 billion and reported a 56 per cent adjusted EBITDA margin. But adjusted operating margin was only 1.0 per cent, and GAAP net loss was US$740 million, with quarterly interest expense of US$536 million. Depreciation on GPUs and debt service consume the economics. CoreWeave buys Nvidia hardware at market prices, finances it with borrowing, and rents capacity into a competitive market. It owns no gate.

The frontier labs invert the popular intuition. OpenAI’s annualised revenue run-rate was roughly US$20 billion at the end of 2025. Anthropic reportedly reached around US$30 billion in April 2026 and about US$47 billion by late May. The growth is real, even if the figures are reported rather than audited. But revenue is not profit. OpenAI’s gross margin has been reported to be around one-third, constrained by inference costs, and internal projections reported publicly pointed to a multibillion-dollar loss in 2026.

Through the Abundance Economics lens, the labs are not toll collectors. They are in demand. A large share of their revenue flows upstream to chips and cloud and lands there as margin. The model itself is becoming less defensible because two forces push price down at once: open-weight models keep closing the capability gap, and the cost of fixed model performance keeps falling. In a layer with low switching costs and credible substitutes, falling input costs cannot be retained. Competition forces it through. That is why a lab can scale revenue explosively and still lose money.

The application layer needs more care. “AI apps” are not one category. Thin wrappers over frontier APIs own no gate and are likely to be crushed. Embedded workflow systems can be different because they control customer data, procurement position, operating processes, or a regulated context. Distribution-owned applications can also hold margin where they already own the user relationship.

Also Read: Give physical AI a soul: Why your voice AI still feels like a bot

Palantir is the clearest example. It is not just “an AI app.” It is an embedded data-and-workflow layer inside government and enterprise operations, and that position can behave like a bottleneck. By contrast, implementation consultants capture demand but earn consulting economics. Accenture may book billions in generative-AI work, but its overall operating margin remains around 15 per cent.

• The first capital-allocation implication is simple: do not price the AI stack as one trade. NVIDIA, TSMC, hyperscalers, neoclouds, model labs, workflow software, and thin apps have different economics because they sit in different places in the pass-through chain.
• The second implication is that revenue is the wrong metric at the model layer. A lab’s run-rate measures how much compute it is buying as well as how much value it is keeping. Treating model revenue like Nvidia revenue is a category error.
• The third implication is that the most durable profit pool may be less glamorous than the market assumes. NVIDIA’s moat is powerful but contestable: the cloud owners’ custom silicon is an attack from above. TSMC’s moat is harder to clone because it rests on physics, capital intensity, yield learning, and years of manufacturing execution. That does not make TSMC risk-free. It has Taiwan exposure, customer concentration, and cyclicality. But the moat itself is structurally harder to erode.
• The fourth implication concerns the capital now being committed. The major hyperscalers are reportedly tracking toward a combined 2026 AI capital spending approaching US$700 billion. That spending is ahead of demand. At the enterprise buyer level, an MIT study found that 95 per cent of organisations deploying generative AI had seen no measurable profit-and-loss impact. If the five per cent of successful deployments scale into the majority, the capital may be repaid. If not, much of the non-bottleneck stack remains a transfer mechanism feeding the gates.

The thesis can break in several ways. NVIDIA’s gate could erode faster than expected if custom silicon scales. A frontier lab could become a true toll collector if one model achieves a durable capability lead, or if regulation entrenches a small number of approved model owners. Enterprise demand could arrive faster than current evidence suggests. Or the binding constraint could migrate from chips to power, shifting the profit pool toward whoever controls dispatchable energy near data centres.

There is also a circularity risk. Some AI demand is financed by the same companies that benefit from it, through equity stakes, cloud commitments, reseller structures, and compute deals. That does not make Nvidia’s or TSMC’s profits fake. Their margins are real. But it does mean some of the revenue feeding those margins may be more fragile than organic demand would be.

The investor question is not whether AI is profitable. It plainly is, at the gates. The question is whether the demand behind those gates is durable enough, and arrives quickly enough, to repay everyone standing in line behind them.

—

The post Is the AI industry profitable? Yes, just not where you’re looking appeared first on e27.