On a Bitcoin 2026 panel about whether data centers can deliver on the AI promise, one speaker put it bluntly: Jensen drove a bus through public cloud, and facilities that used to matter only to enterprise IT or colocated mining suddenly matter to everyone. It is a crude image, and it is also the most honest summary of the last two years I have heard. The accelerator did not just get faster. It reorganized who owns the building.
I keep that line in mind when I read accelerator market forecasts, because the numbers currently in circulation do not agree with each other in interesting ways.
Same market, three different futures
DC Market Insights, in a report added this month covering 2026 through 2035, projects the data center accelerator market reaching USD 605.42 billion by 2035. MarketsandMarkets, publishing in September, valued the market at USD 170.81 billion in 2025 and projects USD 372.68 billion by 2030. Dell’Oro Group, in a forecast published back in February 2025, put AI accelerators at $382 billion by 2029.
Line those up and the disagreement is not about direction. Everyone expects the curve to go up steeply. The disagreement is about when the curve bends.
- Dell’Oro reaches roughly $382 billion a full year earlier than MarketsandMarkets reaches $372.68 billion.
- MarketsandMarkets implies something close to 17 percent annual growth across 2025 to 2030.
- Getting from that 2030 figure to $605.42 billion by 2035 implies roughly 10 percent annually for the following five years.
Those are my rough calculations from the published endpoints, not numbers the firms quoted. But the shape matters more than the decimals: the consensus story is a violent build-out through the end of this decade, then a transition into something that looks more like infrastructure replacement than infrastructure creation. None of the sources I have offer a clean standalone figure for 2026 itself, which is worth saying plainly rather than splitting the difference and pretending it is a forecast.
Why rack-scale changes the unit of analysis
The framing in the DC Market Insights headline is rack-scale AI systems and inference demand. From where I sit as a researcher, those two phrases are the same phenomenon viewed from opposite ends.
Rack-scale means the accelerator is no longer the thing you buy. The coherent memory domain is the thing you buy. When a model’s working set and its KV cache no longer fit in a single device, the interconnect inside the rack stops being plumbing and becomes part of the processor. That is an architectural statement disguised as a procurement decision, and it explains why accelerator market sizing is drifting toward counting whole systems rather than chips.
Inference demand pushes from the other direction. Training spend is lumpy and concentrated in a handful of organizations. Inference spend is diffuse, continuous, and tied to actual usage. And agentic systems change the arithmetic of inference in a way that I do not think the forecasts fully capture yet.
The agent multiplier
A chat turn is one forward pass against a prompt. An agent loop is dozens of forward passes against a context that grows with every tool call, every retrieved document, every intermediate reasoning step it carries forward. Add planning, add self-critique, add retries on failed tool calls, and a single user request fans out into a tree of generations where most tokens are never seen by a human.
That is a different workload profile than the one most capacity planning assumes. It is latency-sensitive at the leaf level but throughput-bound in aggregate. It is memory-bandwidth hungry rather than FLOP hungry. It is bursty in a way that punishes static partitioning. If agent deployment scales the way current adoption curves suggest, the demand driver for accelerators shifts from how many models get trained to how many autonomous loops are running per second.
Follow the cooling money
The funding record from this year makes the physical constraint visible. Orbital Industries closed a $50 million Series B on 28 May 2026 for data center cooling infrastructure. ZutaCore closed a $100 million Series C on 2 June 2026, also cooling. Two raises, five days apart, in the same unglamorous category.
Investors are not funding heat rejection because it is exciting. They are funding it because rack-scale density has made thermal capacity the binding constraint on how much accelerator you can actually install. You cannot buy your way past watts per rack with a purchase order.
What I would watch instead of the headline number
The spread between these forecasts is roughly $233 billion across a six-year window, which should tell you how much model uncertainty sits behind any single figure. More useful signals, in my view:
- Whether memory capacity and bandwidth per rack grow faster than raw compute throughput.
- Whether inference revenue overtakes training revenue in vendor disclosures.
- Whether cooling and power availability start appearing as stated limits in deployment guidance.
The bus went through public cloud. The question now is whether the road ahead is wide enough, and that question gets answered in megawatts and millimeters of coolant tubing, not in press releases.
🕒 Published: