Goldman Sachs backing Spectro Cloud makes sense for one blunt reason: AI infrastructure is expensive, and too much of it still sits idle.
Spectro Cloud is not selling the shiny part of the AI boom. It isn't building the model, making the GPU, or promising you a new chatbot. The Santa Clara company sits in the less glamorous layer underneath, where enterprises have to decide which chips run which workloads, in which cloud, under whose controls.
That layer is getting more valuable. Business Insider reported in October 2025 that Spectro Cloud, a Goldman Sachs Alternatives backed startup then valued at about $750 million, launched PaletteAI with Nvidia to help companies manage AI infrastructure and improve GPU use. CEO and cofounder Tenry Fu told Business Insider that many companies were using only about 30 percent of their GPU capacity, while PaletteAI could push that toward 60 percent.
That's the real story. You can buy the most expensive silicon in the market and still waste the thing you just fought to acquire.
PaletteAI extends Spectro Cloud's original Kubernetes management product, Palette. Cofounder Saad Malik, the company's chief technology officer, described the product to Business Insider as the glue layer that lets different hardware and software components work together. The word glue is plain, but it does the job. Enterprises don't need another dashboard for its own sake. They need the expensive parts of their AI stack to stop sitting around.
The idle GPU problem is not small
Here's the uncomfortable number. Cast AI's 2026 State of Kubernetes Optimization Report, covered by Business Insider and TechRadar, found average enterprise GPU utilization at just 5 percent across the Kubernetes clusters it analyzed. That's not a typo. CPU use was also low, at about 8 percent, according to the same report.
You don't need to be a cloud architect to see the problem. If a company is paying for AI compute and using only a sliver of it, the waste isn't a rounding error. It becomes a board-level cost problem, especially when Nvidia H100s, H200s and newer Blackwell systems remain expensive and hard to plan around.
Much of this behavior comes from the last shortage cycle. In 2023 and 2024, teams learned to grab GPU capacity when they could get it, because waiting meant falling behind. That habit made sense during scarcity. It makes less sense when the bill keeps arriving and the hardware isn't doing enough work.
Recent data center spending shows why this matters now. The Wall Street Journal reported this week that data center owners including Netrality, DataBank, Edged and EdgeCore are exploring stake sales as AI demand pushes infrastructure valuations higher. TechRadar, citing Gartner, reported that AI servers could consume more power than conventional data centers by 2027. The money is moving fast. The power demands are moving with it.
Wasted compute is no longer just a cloud optimization gripe. It's part of the economics of the AI buildout.
Why Goldman would care about orchestration
Goldman Sachs backing a company like Spectro Cloud is not a bet on one chip winning. Frankly, that's the point. Nvidia is still the center of gravity in AI hardware, AMD is fighting for more share, and a long list of inference chip companies are trying to wedge themselves into production workloads. Spectro Cloud can benefit from that fight without having to pick the final winner.
The customer need is practical. A bank, telecom company, airline, defense agency, or cloud provider may want Nvidia systems today and AMD systems tomorrow. It may need workloads in a public cloud, a private data center, and at the edge. It may also have governance rules that make a single-vendor stack impossible. PaletteAI is aimed at that mess.
This is why the Nvidia connection matters, but it shouldn't be misread. Business Insider reported that PaletteAI integrated with Nvidia's AI Enterprise software suite, yet Spectro Cloud pitched the platform as open to other hardware and software vendors. That matters if you're the buyer. Lock-in is easier to sell when budgets are loose. It is much harder to defend when finance starts asking why the GPUs are idle.
For startups, this is a useful reminder. The loudest AI companies are not always the ones closest to the money. Sometimes the better business is in the plumbing: scheduling, utilization, security, cost controls, and all the unglamorous work that makes expensive infrastructure usable.
Spectro Cloud still has to prove that PaletteAI can hold its place as enterprises standardize their AI stacks. Nvidia has its own software ambitions. Cloud providers have every incentive to keep customers inside their own systems. Kubernetes management is also a crowded market, and buyers don't hand over infrastructure control lightly.
But the bet is clear. If companies keep buying AI infrastructure faster than they can use it well, the software that makes that infrastructure behave becomes more important. The chip gets the headline. The orchestration layer gets the invoice.
Also read: InstaLILY raises $60 million to put AI agents inside construction and logistics firms • Shai Morag Raises $60 Million to Build an Identity System for AI Agents • Sequoia Capital Pours $45 Million Into an AI Startup That Calls Its Product an Employee
