Where AI Profit Actually Lands: What the GPU-to-Application Value Chain Means for Your Build-vs-Buy Decision

Enterprise AI budget conversations tend to orbit model costs: which API is cheapest per token, which fine-tuning run fits the quarterly budget. That framing misses the more consequential question, which is where margin actually concentrates across the five-layer stack that sits between a power substation and a deployed AI feature. Public filings from Anthropic and CoreWeave now give technical leaders enough signal to answer that question with reasonable precision, and the answer should change how you structure vendor commitments and internal investment for 2027.

Companion piece to our broader work on AI infrastructure pricing and vendor dependency. See Compute Access Gap: AI Infrastructure Competition for how Anthropic and OpenAI's GPU race is reshaping supply, pricing, and enterprise buyer risk.

The Five-Layer Stack and Where Gross Margin Lives

The AI infrastructure stack runs from physical power through data center real estate, neocloud GPU rental, inference serving, and finally the application layer that enterprise users actually touch. Each layer charges the one above it, and the margin each extracts reflects its structural pricing power, not its engineering complexity.

Power and Data Center

At the base, power and colocation are high-capital, low-margin businesses. Utilities operate under regulated rate structures, and data center operators compete on power availability and cooling density rather than differentiated capability. Gross margins in this tier run in the low-to-mid teens for most operators. The constraint is physical: permitting, grid interconnection, and cooling infrastructure take years to build, which creates scarcity but not pricing power in the traditional sense, because hyperscalers increasingly own this tier directly and bypass the margin entirely.

Neocloud GPU Rental

The neocloud tier, represented most visibly by CoreWeave, sits above colocation and below the model layer. CoreWeave's S-1 filing showed gross margins in the 70-percent range on GPU rental, driven by the spread between wholesale NVIDIA contract pricing and the spot and reserved rates charged to inference providers and enterprise customers. That margin is real but fragile. It depends on sustained GPU scarcity, and as NVIDIA expands H100 and B200 supply, the structural basis for those margins compresses. Burst GPU rates above $13 per GPU-hour reflect current scarcity, not a durable floor.

Inference and Application Layers

Anthropic's disclosed economics place the inference layer at gross margins that, on a per-token basis, remain thin relative to the capital deployed. The application layer is where margin concentration becomes most pronounced. Software companies sitting on top of commodity inference, whether vertical SaaS vendors or internal platform teams, capture the widest spread because their costs are primarily engineering and distribution rather than compute. This is the structural logic behind why application-layer businesses historically trade at higher multiples than infrastructure businesses, and it applies directly to the AI stack.

What This Means for Your Build-vs-Buy Call

The margin distribution tells you something specific about negotiating leverage at each tier. Infrastructure vendors in the neocloud tier currently hold pricing power because GPU supply remains constrained, but that leverage is time-limited. Committing to multi-year reserved GPU contracts at current burst rates locks in costs at a point near the scarcity peak, not the long-run equilibrium.

The application layer presents the inverse dynamic. Because gross margins at the application tier are structurally wide, the vendors selling AI-powered applications to enterprises are extracting substantial margin on top of inference costs that are themselves falling. Building application-layer capability internally, where the underlying inference cost is a commodity input you purchase at market rates, captures that margin spread for your own P&L rather than a vendor's.

The practical implication is not that you should build everything. It is that the application layer is where internal investment has the highest return on margin capture, while the infrastructure layers below it are where you should negotiate hard, commit short, and treat vendor lock-in as a balance sheet risk.

Where Compute Price Increases Will Hit Hardest

As burst GPU rates climb, the pain distributes unevenly across the stack. Teams running inference workloads on spot or on-demand capacity absorb cost increases directly and immediately. Teams with reserved capacity at fixed rates are insulated in the short term but face renegotiation risk at contract renewal if scarcity persists.

The more structurally exposed position is any enterprise that has built its AI architecture around a single neocloud provider without egress flexibility. When compute prices move, the ability to shift workloads between providers is the primary cost lever available to a technical team. Architectures that treat the inference provider as a fixed dependency rather than a substitutable input surrender that lever entirely.

We have written previously about the hidden costs of egress traps and capacity mismatches in GPU deployments. The margin analysis above reinforces that framing: infrastructure vendor lock-in is not just an operational risk, it is a direct tax on the margin you could otherwise retain at the application layer.

Structuring Vendor Commitments for 2027

The 2027 planning cycle should apply a different commitment strategy at each stack layer. At the power and colocation tier, the relevant question is whether your primary cloud or neocloud provider has secured capacity in regions where you need latency compliance, because that is increasingly a supply constraint rather than a price negotiation.

At the neocloud tier, the appropriate posture is shorter commitment windows than vendors will prefer to sell you. GPU scarcity is real today and will be less acute by late 2027 as supply expands. Accepting a two-year reserved commitment at current rates to secure a modest discount is a bet that scarcity persists, and the supply trajectory does not support that bet with high confidence.

At the inference and application tiers, the investment logic reverses. Longer-term internal capability building at the application layer compounds in value as inference costs fall, because your fixed cost base shrinks while your application margin expands. The teams and tooling you build internally in 2026 and 2027 are not exposed to GPU price volatility in the same way your infrastructure spend is.

Reading the Stack Before You Sign

The practical use of the margin analysis is not academic. Before committing to any significant AI infrastructure contract, map the gross margin profile of the vendor you are negotiating with and ask what gives them pricing power in their tier. If the answer is GPU scarcity, the leverage is temporary. If the answer is proprietary data, network effects, or switching costs at the application layer, the leverage is durable and you should price that into your negotiation posture accordingly.

The visibility that Anthropic and CoreWeave's filings now provide is genuinely useful for enterprise buyers. It is the kind of information that, historically, only sophisticated infrastructure investors had access to. Technical leaders who use it to inform vendor selection and internal investment allocation will make materially better capital allocation decisions than those who continue to treat AI spend as an undifferentiated cost line.