The narrative around AI is almost entirely about what it can do for you. The conversation about what it costs — and who designed that cost — is much quieter. This is the second conversation, end to end: the trillion-dollar bet being built above you, and the bill that lands on you — as an individual, as an enterprise, and as the startup that went 100% autonomous before anyone modelled the meter.
- The companies building AI have committed hundreds of billions to infrastructure — and their revenue model is your consumption, priced like a utility.
- Chips and power are the two real constraints: HBM memory is now ~63% of AI chip cost, and the industry is signing 20-year nuclear contracts to keep the meters running.
- The bill at every level, with receipts: my ~€1,000/month personal stack, the $500K–$1M+/year mid-size enterprise, and the startup paying ~$500K/month for 5–10 autonomous agents.
AI promises to make you more productive. That promise, broadly, is real. The question being obscured is: productive on whose infrastructure, at whose pricing, under whose model of consumption? The companies building AI are not charities. They are making the largest infrastructure bets in modern technological history — and they need a revenue model that justifies those bets.
That revenue model is you. And me. And every enterprise paying per seat, per token, per API call.
What Is Actually Being Built
In January 2025, the Stargate project was announced: a joint venture between OpenAI, SoftBank, Oracle, and MGX, with $500 billion committed over four years to AI infrastructure in the United States. Not research. Not product development. Infrastructure — data centres, compute clusters, power grids.
The opening deployment was $100 billion, beginning immediately in Texas. The message from the announcement was unambiguous: we are building the pipes through which intelligence will flow, the same way AT&T built telephone infrastructure and Cisco built the early internet backbone. The bet is that those pipes will be essential, permanent, and profitable.
These are not venture bets on speculative technology. These are infrastructure bets on a utility model — and utility models have one defining characteristic: the infrastructure owner wins when consumption grows.
Where does this curve go if it holds? The most rigorous public attempt to answer that is the AI 2027 scenario by the AI Futures Project — a research-grade forecast, written by former OpenAI and AI-policy researchers, that models the infrastructure race quarter by quarter. Their projection for late 2026:
Treat those numbers as what they are — a scenario, not a measurement. But note the direction: every verified figure in the grid above is a committed step toward them.
Chips and Power — The Two Things the World Does Not Have Enough Of
Building AI infrastructure at this scale requires two inputs: specialised compute and electricity. Both are under severe constraint.
On compute: NVIDIA controls roughly 80% of the market for AI training and inference hardware. An H100 GPU — the workhorse of large model training — peaked at $25,000 to $40,000 per unit in 2023-2024. NVIDIA’s data centre revenue grew from $47 billion in FY2024 to $115 billion in FY2025. Demand has consistently outpaced supply. TSMC, which manufactures NVIDIA’s chips, is racing to add capacity in new fabs across Taiwan and Arizona. They are not building fast enough.
And inside each chip, the constraint has a name: memory. Epoch AI’s component cost analysis of the chips designed by NVIDIA, AMD, Google, and Amazon shows high-bandwidth memory (HBM) growing from 52% to 63% of total chip cost between early 2024 and late 2025, while every other component’s share shrank. In absolute terms, HBM spend across those four designers nearly tripled — from roughly $12 billion in 2024 to $32 billion in 2025. The bottleneck is already repricing the hyperscalers’ plans: Microsoft raised its 2026 capex outlook by $25 billion citing component costs, and Meta raised its range by $10 billion for the same reason. When one component of one part of the supply chain moves two of the world’s largest companies to revise guidance by tens of billions, that is not a procurement detail. That is the meter being built into the silicon.
On power: the International Energy Agency (IEA) projected in 2024 that data centre and AI electricity demand would double by 2026 — from roughly 400 terawatt-hours to over 800 TWh annually. For context: 800 TWh is approximately the annual electricity consumption of Japan. The industry’s response to this has been, unusually, nuclear:
When the companies building AI are signing 20-year nuclear power contracts, they are not speculating. They are committing to a consumption curve that requires energy infrastructure that doesn’t yet fully exist. That commitment is priced into what you pay per API call.
The People Running This Are Not Subtle About the Roadmap
The strategic framing from the people at the top of this industry is worth reading carefully — not for what they say AI will do, but for what they say AI will cost and how it will be consumed.
These statements are not made in isolation. They come from the CEOs of companies that have committed, collectively, hundreds of billions to AI infrastructure. The framing — that AI is essential, inevitable, and will be priced like a utility — is not a neutral market observation. It is the monetisation roadmap of the infrastructure they are building. Their return on investment depends on your consumption growing.
So that is the wager above you. Now the other half: what it costs to operate inside it. I will start with my own numbers, because I think the most honest version of this conversation is one where the person giving advice is transparent about what they actually spend.
My Stack, My Bill
I track my AI spend carefully. Not because I resent it — I think the value is real — but because I believe the only honest relationship with any tool is knowing exactly what you are paying for it. Here is what my personal stack costs each month, as of early 2026:
That is roughly what a junior developer earns per month in some European markets, spent entirely on AI tool subscriptions. Most of my active usage — around 20 hours per week — happens outside working hours: personal projects, architecture thinking, writing, learning. I do this by choice, and the productivity return justifies it for where I am in my career. But I say that clearly: this is a real cost with a real payoff that requires honest accounting. It is not a hobby expense. It is an infrastructure investment in my own capability.
The question worth asking before you build a similar stack: what would this cost look like if the tools stopped working the way they do today? What if the model that underpins your main productivity tool gets rate-limited, deprecated, or repriced? These are not hypothetical risks. They are scheduled events in an industry that is still figuring out its sustainable pricing model.
What the Enterprise Bill Actually Looks Like
At the enterprise level, the bill compounds differently. The individual subscription becomes a seat count. The occasional API call becomes a committed monthly consumption tier. And the procurement conversation that starts with "we're just adding Copilot" tends to end somewhere considerably more expensive.
Let me use GitHub Copilot as a specific example, because it is the most widely deployed AI developer tool in enterprise settings and its pricing structure is unusually instructive about how enterprise AI cost actually works. One honesty note before the table: the first three rows are GitHub's published list prices. The three extension bands below them are not public pricing — they are composite figures from enterprise rollouts I have seen, included because the base subscription is never where the bill ends:
| GitHub Copilot tier | Monthly cost | What it covers |
|---|---|---|
| Individual | $10 / user | Basic completions, chat in IDE, standard models |
| Business | $19 / user | Policy controls, audit logs, IP indemnity, org management |
| Enterprise | $39 / user | Custom models, knowledge bases, fine-tuning, advanced agents |
| Extensions (Standard) * | $25–60 / user Per-user monthly quota on top of base subscription | Third-party integrations: Jira, Datadog, Sentry, etc. |
| Extensions (Premium) * | $100–500 / user Elevated access tiers for power users | Advanced agent capabilities, higher model access quotas |
| Extensions (Enterprise-grade) * | Up to $1,000 / user Requires Business Unit Director approvalApproval gate | Custom agent workflows, dedicated model capacity, full API access |
That approval gate at $500+ per user per month is not a product limitation. It is a governance control that enterprises are having to build in real time, because no one designed their procurement process for a world where a developer productivity tool could cost $12,000 per person per year before the base subscription. The control exists because someone, somewhere, approved $500/month per user across a large team and the quarterly invoice produced a conversation that nobody wanted to have.
The procurement conversation that starts with "we're just adding Copilot" tends to end somewhere else. The initial subscription is visible and approved. What follows is a series of smaller approvals — a specialist AI extension here, an API integration there, a premium agent tier for the senior developers — that individually seem reasonable and collectively exceed the original budget by a margin that surprises nobody in retrospect.
A mid-sized technology company deploying GitHub Copilot Business ($19/user) across 200 developers, adding Microsoft 365 Copilot ($30/user) for the broader organisation at 500 seats, plus Azure OpenAI for three internal integrations, is at $500,000 to $700,000 annually before any additional AI tool spending. Add data, marketing, and operations AI tools, and you are at $1M+ per year routinely. The line items were all reasonable. The total required a board conversation.
None of this criticises the tools themselves — the value is, in many cases, real. The point is how utility models expand. You do not decide once to spend X on electricity. You decide to add each appliance, and the bill follows behaviour. Enterprise AI billing is heading exactly this direction — and the companies designing these pricing models know it.
What Happens When You Go All-In on Autonomous Agents
The most instructive cost story in AI right now is not the individual developer or the cautious enterprise. It is the startup that decided to take the pitch seriously and build with 100% AI development from day one.
This is a pattern happening at multiple companies, in different forms. The trajectory looks like this:
This is a real description of a real team. The productivity numbers are real. The outcome — serving 1M+ users with a fraction of the traditional engineering headcount — is genuinely remarkable. The investors are satisfied. The metrics are impressive.
What is not in that description: the monthly AI bill. Running 5–10 autonomous coding agents simultaneously, at multi-step execution depth, with full RAG retrieval and MCP tool calls at each step, is an infrastructure cost that scales with every agent task, every context refresh, every tool call, every review cycle — a different category of spend from a $100/month tool subscription. At this level of usage, the monthly AI bill for a small engineering team routinely reaches $500,000 per month or more.
Investors are currently funding this. They are satisfied because the output metrics — features shipped, users served, engineering velocity — look extraordinary. The structural question being deferred: what happens to the AI bill when the application scales further and the agents start encountering the parts of the codebase that don't have clean bounded contexts, clear inputs and outputs, and predictable failure modes? The answer is that the agent execution depth increases, the retry rate goes up, the cost per feature climbs, and the productivity numbers that justified the model start to compress.
The caution here applies to operating any powerful tool without understanding its cost model at scale — AI is simply the current example. The startup that goes from "AI as a code assistant" to "10 autonomous agents in parallel" in two months has made a bet. The bet may pay off. But the cost structure of that bet is real, and the deferred question — what does this look like at ten times the current scale? — is the one that will eventually produce a difficult board conversation.
The Cost Nobody Puts on a Spreadsheet
The dollar cost of AI is the visible cost. There is a second cost that compounds alongside it, and it is less visible because it does not appear as a line item: the quality degradation that emerges when AI adoption outpaces the process design that should contain it.
What high-percentage AI adoption looks like six months in
- Context is not shared. Each developer's AI session has different history, different loaded files, a different mental model of the system. There is no collective memory. The team's understanding of the codebase is distributed; the AI's understanding is fragmented per session.
- Codebases exceed context windows. A large production system cannot be held in any model's context simultaneously. Every AI interaction is working with a partial view. The larger the system, the larger the blind spot — and the more the agent's confident output diverges from the system's actual state.
- Hallucination rates multiply with complexity. A 2–4% incorrect suggestion rate that seems acceptable in isolation compounds across a team. Five developers each accepting one wrong suggestion per day is twenty incorrect pieces of code before end of sprint — many of which pass automated tests because the tests were also generated by AI, against the same incomplete context.
- The review burden shifts but does not disappear. The senior engineer's role becomes less about writing code and more about detecting subtle errors in AI-generated code — a different skill, and in some ways a harder one, because the errors are less obvious and more architecturally coherent.
The solution to this is not less AI. It is process design that accounts for how AI actually works — its context limitations, its tendency to be confidently wrong, and its inability to hold shared state across sessions. The dollar cost of AI is knowable in advance. The process cost of deploying it wrong at scale is not — and it compounds in ways that a subscription renewal notice will not warn you about.
The meter is running for every individual developer, every engineering team, and every startup directing autonomous agents at a codebase. The question is not whether to pay — it is whether you know exactly what you are paying, why, and what the bill looks like at two times your current scale.
None of this is an argument against using AI. I use it constantly, and my honest assessment is that it is the most significant productivity amplifier I have worked with in more than two decades of professional engineering. The value is real.
What I am arguing for is intentionality. The companies building the infrastructure win when consumption grows — and growth is their goal, not optimisation of your bill. Your electricity provider does not nudge you toward energy efficiency. Your AI provider does not nudge you toward using fewer tokens. Budget for AI the way you budget for cloud infrastructure — with actual numbers, reviewed quarterly, with ceiling thresholds and an honest assessment of what the consumption is producing.