The most useful number in the data center accelerator forecasts isn’t any of the headline figures. It’s the gap between them. And if you build bots for a living, that gap should change how you architect, not how you feel about the AI boom.
Here’s what landed recently. DC Market Insights put out a report covering 2026 through 2035, projecting the data center accelerator market at USD 605.42 billion by 2035. MarketsandMarkets pegged the market at USD 170.81 billion in 2025 and forecast USD 372.68 billion by 2030. Dell’Oro Group, back in February 2025, called for $382 billion by 2029.
Line those up and something odd shows up. Dell’Oro sees $382 billion arriving in 2029. MarketsandMarkets sees roughly the same figure a year later, in 2030. Then DC Market Insights adds another 230-ish billion over the following five years. Three firms, three methodologies, three different shapes for the same curve. None of them, notably, give a clean answer for what 2026 itself looks like.
Forecast Spread Is Information, Not Noise
I’ve stopped reading these reports for their endpoints. A ten-year projection is a bet on compounding assumptions: power availability, fab capacity, how fast inference workloads grow, whether rack-scale systems get adopted at the pace vendors hope. Small changes in those inputs produce wildly different 2035s.
What the spread tells me is that nobody has a confident read on the back half of the decade. Analysts agree the near term is steep. They diverge the moment the forecast depends on second-order effects. That’s a reasonable thing to admit, and it’s more honest than a single tidy number.
For builders, the practical read is this: capacity will keep expanding, but you should not plan your cost model around the assumption that expansion makes accelerators cheap for you specifically. Growth in market value means more money moving through the space. It doesn’t mean your per-token inference cost drops on a schedule you can bank on.
Follow The Cooling Money
The funding news from mid-2026 caught my attention more than the forecasts did. Orbital Industries raised $50 million in a Series B on May 28, 2026, for data center cooling infrastructure. ZutaCore pulled $100 million in a Series C on June 2, 2026, also cooling.
Two meaningful rounds in a week, both for heat removal. That’s a thermal story, not a chip story. Rack-scale AI systems concentrate enormous power density into a single rack, and air stops being adequate well before the silicon stops being capable. Investors funding cooling are effectively betting that the constraint on deployment is physics and facilities, not demand.
If you’ve ever had a deployment queue sit idle waiting for capacity in a specific region, you’ve felt the downstream version of that constraint. Accelerator availability is increasingly a question of which buildings can actually keep them running.
What Rack-Scale Changes For Bot Architecture
Rack-scale systems treat a whole rack as one coherent compute unit with very fast interconnect across it. That’s great for enormous models. It’s also a reminder that the infrastructure is being optimized for a class of workload most bot builders don’t run.
Most of what I build is inference, and most of that inference is small, bursty, and latency-sensitive. A support bot answering one question. A tool-calling agent making three sequential API decisions. None of that needs a rack. It needs predictable latency and a cost per call that doesn’t ruin unit economics.
So my working rules, given all of the above:
- Treat the model as a swappable component. Route through an abstraction so you can move between providers and model sizes without rewriting your agent logic. Pricing and availability shift faster than your codebase should.
- Measure cost per completed task, not per token. A cheaper model that needs three retries and two extra tool calls is not cheaper.
- Push work off the accelerator where you can. Caching, retrieval, deterministic validators, and plain code handle more than people expect. Every call you don’t make is a call you don’t pay for or wait on.
- Design for regional capacity gaps. If cooling and power are the real limits, assume some regions will be tight. Fallback routing is cheap insurance.
- Don’t over-provision against a forecast. Analysts disagree by hundreds of billions. Your capacity planning should be shorter-horizon than theirs.
Where That Leaves Us
The accelerator buildout is real and large, and three credible firms expect it to keep growing through at least 2035. That’s the solid part. The precise trajectory is contested, and the money flowing into cooling startups suggests the bottleneck sits in facilities rather than appetite.
Build like inference stays expensive and capacity stays uneven. If the cheerful forecasts prove right, you’ll be pleasantly surprised. If they don’t, your bot still runs.
🕒 Published: