Between 30 and 60 percent of the GPUs that enterprises reserve sit idle at any given moment. Demand has not fallen. The capacity is stranded in silos — one team's allocation going cold while another team waits in queue for the same silicon. By the industry's own arithmetic, that stranded compute costs more than $240 billion a year.
A four-person startup in Seattle thinks the fix is software. Chamber, founded in 2026 and accepted into Y Combinator's Winter 2026 batch, sells an agentic platform that finds idle GPUs and puts them back to work. The company says it can run roughly 50 percent more workloads on the same machines without buying a single new card.
Stranded enterprise GPU capacity
Share of reserved GPU capacity sitting idle in siloed allocations · Chamber / Y Combinator, 2026
Utilization dashboards are not new. Every cloud ships a scheduler, and every platform team keeps a spreadsheet tracking who reserved what. Its bet is that the hard part was never measurement.
The GPU nobody is using
The software watches a fleet across AWS, Google Cloud, Azure, and on-premise racks. It detects unhealthy nodes, reschedules failed training jobs from their last checkpoint, and moves work onto capacity another team has left cold. A two-phase scheduler keeps each team's reserved pool intact while letting jobs burst onto elastic headroom.
The output is a decision, not a chart. The agent right-sizes allocations, reorders the queue by priority, and pushes alerts into Slack, email, and PagerDuty. Done well, that lifts utilization enough to absorb about half again as many workloads on hardware already paid for. The company holds SOC 2 Type I and Type II certification — an independent security audit that enterprise buyers require before granting an agent access to their clusters.
From FLOPS purchased to FLOPS actually used.
A four-person team with hyperscaler scars
The founders built this kind of system before, at a scale that matters. Andreas Bloomquist led the launch of Amazon's central GPU orchestration service and AWS's CloudWatch Application Signals. Shaocheng Wang spent nine years shipping storage and observability products at AWS. Charles Ding ran infrastructure teams at Meta and Amazon, and this is his second company after one exit. Jason Ong came from Amazon and Flexport.
Chamber has four employees and a seed check of $500,000 from Y Combinator. That is the whole balance sheet.
The credibility is the résumé. The company is selling the same problem it once solved internally at a hyperscaler — and the buyers it targets are the ones who watched that internal system work.
Orchestration, not more silicon
Timing helps. AI capital spending has moved from training runs toward inference, where one GPU serves many small requests and idle minutes are pure loss. When compute was scarce, nobody cared about a wasted node. Now the largest buyers are signing multi-year commitments measured in billions, and utilization is the line that decides whether those deals pay off.
That reframes the hardware race. Buying more cards is easy and expensive. Running the cards you already own closer to capacity is cheaper and harder — which is exactly the gap a small team can attack while the incumbents sell more boxes.
The seed round is the question mark
$500,000 does not buy a sales team. The category is crowded: observability vendors, cloud schedulers, and Nvidia's own tooling each touch a piece of the same job. If orchestration turns out to be a feature rather than a platform, an incumbent absorbs it and the startup disappears into someone else's product page.
This is the latest in a run of AI-infrastructure stories we have tracked this month where the binding constraint turned out to be operational — wiring, power, and now idle silicon.
The real edge is pedigree — its founders built Amazon's internal GPU orchestration — not a novel algorithm.
The open risk is scale: a $500,000 seed against deep-pocketed incumbents, in a category that may collapse into a feature.