Created At: Tue Jun 09 2026
The AI Infrastructure Bet
A Thought Experiment on the Success Parameters
Nobody knows how the gigantic IPOs on agentic AI and their infrastructure will end. This article is not a hedge — it is trying to be an honest starting position when the hype is this large, and the timelines this compressed.
What needs to be done, is laying out the parameters. Map what has to be true for the bet to pay off. Map what has to be true for it not to. And ask, whether we have seen this particular shape of optimism before.
I guess we have.
The Scale of the Commitment
In 2024 and 2025, Microsoft, Google, Meta, and Amazon collectively committed north of $300 billion in AI infrastructure capital expenditure. Data centres, GPU clusters, cooling systems, power infrastructure, long-term energy contracts. These are not software bets. They are physical assets with 10–15 year depreciation schedules and energy obligations that do not renegotiate when the market moves.
Nvidia collected most of the hardware revenue. Their data centre division generated $47.5 billion in fiscal year 2025 — a number that would have been implausible to forecast three years earlier. They sold the equipment. They do not carry the operational risk of what it produces.
OpenAI, the company most publicly associated with the AI moment, is burning approximately $5 billion per year on compute and operations against a valuation that peaked above $150 billion. That valuation is not supported by current revenue. It is supported by a projection — that frontier model capability will remain scarce, that enterprises will pay premium pricing for access to it, and that the switching costs will be high enough to make those relationships durable.
Each of those assumptions is load-bearing. Remove one and the math changes materially.
What Has to Be True for the Bet to Pay Off
The success scenario requires several things to hold simultaneously, over a long enough window for the infrastructure to depreciate against real revenue.
Frontier models must remain meaningfully better than open-source alternatives for commercially critical tasks. If the capability gap closes — and the trajectory of Llama, Qwen, Mistral, and their successors suggests it is closing — then enterprises will migrate toward local deployment. The per-token economics of running a capable open-source model on owned infrastructure are not comparable to API pricing at scale. Enterprise procurement departments will find this out. They always do.
The application layer must not capture all the margin. Microsoft's Copilot embedding strategy, Salesforce's Einstein layer, ServiceNow's AI additions — these are all application-layer plays that use model APIs as an input cost and charge customers for the workflow, not the inference. If that pattern dominates, the model providers become commodity suppliers to application vendors. Commodity suppliers do not generate the returns that justify $300 billion in capital commitment.
New use cases must emerge at a rate that expands the total addressable market faster than pricing compresses. The argument here is that falling inference costs unlock use cases that don't exist yet — that at $0.001 per token, entirely new categories of automated work become viable. This is plausible. It is also the standard argument made by every infrastructure provider facing margin compression. Volume compensates for price decline — until it doesn't.
Genuine AGI-adjacent capability must arrive before open-source catches the current frontier. This is the deepest bet. If a qualitative capability leap occurs — something that makes the current generation look like a calculator looks to a modern processor — then the infrastructure advantage becomes structural and the open-source parity argument collapses. The organisations with the most compute win decisively and permanently. This is the internal narrative driving the investment. It may be correct.
What Has to Be True for It to Fail
The failure scenario does not require a dramatic event. It requires the success conditions above to drift, gradually, in the wrong direction.
Open-source model capability closes to within a commercially irrelevant margin of frontier performance on the tasks enterprises actually run — document processing, support automation, content generation, code assistance. This is already partially true today. Llama 3.1 405B competes with GPT-4 class on most standard benchmarks. The gap that existed 18 months ago was categorical. Today it is a margin. Margins compress.
The hardware depreciates on schedule. GPU generations turn over every two to three years. The H100 clusters being installed today will be architecturally outdated before the infrastructure debt they were purchased to service is retired. The companies that bought the hardware carry that depreciation regardless of what the model market does.
Enterprise procurement, once it understands the local deployment economics, executes a slow migration away from API dependency. Not all at once. Gradually, starting with the highest-volume, lowest-complexity workloads — exactly the category where local OSS models are already sufficient. The API revenue base shrinks from the bottom up, which is how commodity markets always compress.
The qualitative capability leap does not arrive on the timeline the investment assumes. The scaling hypothesis — more compute, more data, emergent capability — has delivered diminishing returns at each successive generation. The jump from GPT-3 to GPT-4 was smaller than the jump from GPT-2 to GPT-3, despite a reported order-of-magnitude increase in compute. There are physical limits to what parallel GPU computation can achieve. Signal propagation between chips, coordination overhead at cluster scale, the energy cost of moving data across interconnects — these are not software problems. They are physics problems.
The Shape We Have Seen Before
In 1999, the global telecommunications industry was building fibre optic networks at a pace the projected demand for internet traffic appeared to justify. Analysts modelled exponential growth in bandwidth consumption. The infrastructure investment followed. Companies like WorldCom, Global Crossing, and 360networks raised and spent billions on undersea cables, terrestrial fibre, and switching infrastructure.
The demand projections were wrong — not directionally, but on timing. Internet traffic did grow exponentially. It just did not grow fast enough, soon enough, to service the debt the infrastructure companies had taken on. They went bankrupt. The assets were acquired at cents on the dollar by the survivors. The infrastructure itself — the fibre in the ground, the cables under the ocean — eventually carried the traffic it was built for. The companies that paid for it were gone.
Cisco, the equipment vendor, collected its revenue throughout. It sold the routers and switches. It did not carry the operational risk of the networks those devices were installed in. Sounds familiar?
The pattern is not that the infrastructure was wrong. The fibre was used. The pattern is the mismatch between the capital commitment timeline and the revenue realisation timeline — and who bears the risk of that mismatch.
The Open Question
The current infrastructure bet may pay off. The qualitative leap may arrive. Frontier pricing power may prove more durable than the open-source trajectory suggests. New use cases may expand the market faster than margins compress. History does not repeat mechanically, and there are real differences between 1999 and now — the companies making the infrastructure bet today have existing revenue bases, diversified businesses, and balance sheets that WorldCom did not.
But the shape of the argument is familiar. Large capital commitments made against projected demand that does not yet exist. Equipment vendors collecting revenue with no exposure to the outcome. Application-layer companies capturing the margin that is supposed to service the infrastructure debt. And a core assumption — that the technology will produce something categorically new before the economics demand it — is doing enormous work without being named as the thing it is: A hypothesis.