Remember when GPUs were a gaming thing? When the biggest question about a graphics card was whether it could hold 144fps in whatever shooter was popular that year? That era ended quietly, and now the same silicon sits in racks that cost more than houses, and press releases about it arrive in my inbox at a rate of roughly three per week.
This week’s batch is a good example of why I’ve gotten skeptical about market sizing reports in general. DC Market Insights just put out a forecast saying the data center accelerator market will hit USD 605.42 billion by 2035. That’s a big number with a decimal point in it, which is always a nice touch. Precision to the tens of millions, a decade out.
The numbers don’t agree, and that’s the interesting part
Here’s what makes this worth looking at rather than just forwarding along. MarketsandMarkets has the same market at USD 372.68 billion by 2030, starting from USD 170.81 billion in 2025. Dell’Oro Group forecasts $382 billion by 2029. Three firms, three methodologies, three numbers that are close enough to look like consensus and far enough apart to mean something very different depending on which one you’re planning around.
Dell’Oro and MarketsandMarkets land in roughly the same territory, just a year apart. Fine. But DC Market Insights is projecting another 220-plus billion on top of that over the following five or six years. That’s not a rounding difference. That’s a different story about what happens after the current build-out wave crests.
And none of these sources gives a clear answer for what 2026 actually looks like. Which is funny, because 2026 is the year we’re in.
What this means if you’re actually buying tools
I review AI toolkits. People send me things, I try them, I tell you whether they work. So my interest in a $605 billion forecast is pretty narrow: does it tell me anything about what I’ll be able to run next year, and what it’ll cost?
Mostly, no. But the forecast framing does flag two things that show up in my actual testing:
- Rack-scale systems are the unit now. The DC Market Insights report ties its growth to rack-scale AI systems rather than individual chips. That shift has downstream effects on what you can rent by the hour and in what shapes. If the supply side is organized around full racks, the cheap single-GPU tier gets weirder.
- Inference demand is driving deployment. This matches what I see in pricing for the hosted inference services I test. Training capacity gets the headlines. Inference capacity is what determines whether your app’s latency is acceptable at 2pm on a Tuesday.
The funding activity around this is mostly in cooling, which tells you where the bottleneck is. Orbital Industries raised $50M in a Series B in May 2026 for data center cooling infrastructure. ZutaCore pulled $100M in a Series C in June. When the money goes toward getting heat out of buildings, that’s a hint that the limiting factor isn’t chip design anymore.
My honest read
I don’t think anyone knows what this market looks like in 2035. I don’t think the firms publishing these numbers would claim otherwise if you asked them directly over a beer. Ten-year forecasts in a space where the dominant architecture changed twice in the last five years are directional at best.
What I’d actually use these reports for is the near-term shape. A market going from $170.81 billion to around $375 billion in five years is a solid, high-growth trajectory that’s already visible in hardware availability and rental pricing. You can plan around that. You can make decisions about whether to buy or rent, whether to commit to a provider, whether to build your stack around assumptions of abundant or scarce compute.
The 2035 number is a conversation piece. It’s useful for pitch decks and board slides and justifying capital expenditure to people who want a big round figure on a chart.
If you’re picking tools in the next twelve months, the practical takeaway is simpler. Inference capacity is where the demand is going, cooling is where the constraint lives, and the rack is becoming the unit of purchase. Those three facts will shape your options more than any decimal point in a 2035 projection.
Build for the capacity you can verify exists. Treat the forecasts as weather, not as a schedule.
🕒 Published: