- Cerebral Valley
- Posts
- GPU clouds everywhere are sold out: Shadeform is building the solution.
GPU clouds everywhere are sold out: Shadeform is building the solution.
Founder/CEO Ed Goode on why AI startups need to think like asset-based businesses.
Cerebral Valley
May 14, 2026
CV Deep Dive
Shadeform is the GPU cloud marketplace betting that the way through the compute capacity crunch is multi-cloud, not bigger hyperscalers. Founded in 2023 by Ed Goode and Ronald Ding (YC S23), Shadeform unifies GPU supply across 30+ cloud providers - including Nebius, DigitalOcean, Lambda, and Crusoe - under a single API and console. Teams get instant access to thousands of on-demand and reserved GPUs without quotas, account management, or vendor lock-in.
Since their first CV Deep Dive in February 2025, Shadeform has scaled materially: customer mix has shifted from AI-native startups to also include hyperscalers, Fortune 500 enterprises, leading inference providers, and NVIDIA themselves. Today, the company is announcing Shadeform Capacity Solutions - a new offering that expands compute options beyond clouds’ existing infrastructure. From data center sourcing through hardware procurement to cloud operator management, Shadeform now packages and delivers entire managed clusters for teams that need predictable capacity at scale.
In this conversation, Ed walks us through what's actually happening in the GPU market right now, why software businesses are being forced to think like asset-based businesses for the first time, and what Capacity Solutions unlocks for teams trying to keep pace with the demand.
Let’s dive in ⚡️
Read time: 8 mins
Our Chat with ED 💬
Ed, welcome back to Cerebral Valley. We last spoke in February 2025 - quite a bit has changed in the GPU market since then. Walk us through what you're seeing right now from your seat at Shadeform. What does the supply picture actually look like today?
It's a completely different world. Back in February 2025, whether you were an individual, startup, or large enterprise, you could generally get the GPU capacity you needed on-demand. Our marketplace of 30+ providers made that seamless.
Now it’s a different story. New always-on agentic workloads combined with accelerating enterprise AI adoption has driven a dramatic increase in GPU consumption. The pool of available capacity has become so small that GPU providers are pulling their short term offerings just to keep up with AI lab demand.
What we’re seeing now is that pretty much everyone, the large labs, enterprises, and every size startup, is struggling to find even reserved capacity. Most are being given three- to six-month long timelines, and some are being turned away entirely.
Microsoft is reportedly sitting on $80B in unfulfilled Azure orders. Hyperscaler capex is projected at $660B+ for 2026, but roughly half of planned U.S. AI data centers have been delayed or canceled. From where you sit, which is the harder constraint right now - chips, power, or something else?
Candidly, it's all of the above. You can rank it hierarchically. The biggest constraint in 2026 is available colo and data center space. It's practically tapped out. You can squeeze out a megawatt here or there, but even that is starting to thin out.
Then there are CPU bottlenecks, GPU bottlenecks, and a ton of switch and networking bottlenecks. But anything at scale is extremely limited from these large projects. As you can see in the reports and timelines, projects are even getting canceled and delayed because they're running into actual IT and power infrastructure supply bottlenecks. Even commodity minerals are running into these problems.
So it's not a single point - but the immediate constraint that's going to prohibit a team from moving forward with their scale right now is data center capacity.
How is this showing up in customer behavior on Shadeform today? Are you seeing teams scrambling, locking in long-term reservations, or behaving differently than they were six or twelve months ago?
Definitely. In the mid-stage AI-native startup space, where teams were comfortably reserving on monthly, three-month, or six-month terms a year ago, they've had to become comfortable with reserving for two- to three-year contracts. It's just a competitive market for inventory, and if you're looking for something flexible, you're at somewhat of a disadvantage now.
Fortunately, we've been working with some very accredited partners to bring shorter-term capacity to Shadeform. But as we put those online, they get booked out relatively quickly.
What’s been most revealing is hearing how our new customers had been navigating this shortage before finding Shadeform - they all generally had the same experience. Most had been working with one or two clouds, then hit a roadblock where those clouds couldn’t get them the capacity they needed on a fast enough timeline. The average we’re hearing for most chips is three- to six-months, sometimes twelve.
From there, they end up reaching out to every provider they can find until one tells them they can deliver. But unfortunately, in many cases, providers will promise capacity on the optimistic bet that they can get something deployed in time, then hit their own roadblocks further down the supply chain. In more extreme cases, we’ve heard of providers pulling capacity out from under a customer right before they sign in favor of a better deal.
It’s a wild market, and a big part of the pain for customers seems to be a lack of visibility and foresight. They don’t plan to scale until something starts to break, and they don’t know where capacity is, so they end up burning a significant amount of their time trying to track it down at the last minute, often with very frustrating results.
What does this mean for the AI startups and engineers reading this newsletter? If you're a Series A or B company building inference today, what should you be doing differently in the next 90 days?
There’s a fundamental shift happening. A year ago, teams could get away with not paying attention to the GPU supply chain. There was a general assumption that, when you needed capacity, it would be available.
That’s not the case anymore. AI teams are having to think like traditional asset-based business, forecasting their demand and planning months to years out to guarantee resources.
Startups need to get serious about logistics, find ways to increase visibility, and familiarize themselves with this new landscape.
If you aren't thinking about your startup as an asset-based business and managing your resources with that paradigm shift, you're going to be caught in a tough position.
Introducing Shadeform Capacity Solutions
That brings us to your new offering - Shadeform Capacity Solutions. Give us the elevator version: what is it, and who is it for?
Given the difficulties founders face adapting to this new reality, we take the overhead of managing a compute asset-based business off your plate so you can focus on the actual end product, not all the infrastructure and logistics that deliver it.
We start with visibility. Over the years, we’ve developed a deep partner network with almost every major cloud, data center, colocation provider, and hardware OEM. We’re able to leverage those partnerships to give our customers a complete market overview of not just all available GPU capacity and upcoming deployments, but also custom build-out options based on hardware and data center availability. Without visibility, you can’t plan effectively, and this is the most complete picture you can get.
From there, our solutions architects work with your engineering team to understand your demand picture and technical requirements, fitting you with the right capacity for your unique needs. We’ve been building a highly competent team with exceptional talent from top clouds like Coreweave, Lambda, Azure, and AWS
Then we move. If the right fit ends up being existing deployments from our partners, we deploy those clusters for you in our unified platform with centralized billing, monitoring and support.
If there isn’t existing capacity available, then we build it. What that entails for Shadeform is that we go out and work with the data centers, colo providers, and colo brokers to get the necessary data center space for you. We’ve built great relationships with many of them and often get early visibility into capacity before it hits the open market. Then we work directly with the OEMs - Supermicro, Dell, HP, Lenovo - to source and provide the compute along with all the networking infrastructure needed to run a cluster. Finally we bring in a dedicated cloud operator to manage that cluster.
The way to think about it is that we're essentially the developer of these assets - we put them together, then sell and deliver them to our cloud partners to manage and operate. In a real estate model, our partners are the owner and property manager; we're the developer.
So we run the whole process to get you a managed cluster under the care of a quality operator. If you want to manage it yourself, we also deliver it to end customers who want to actually own the asset, and we have a host of solutions there to reduce TCO. But if you want it cloud-operated, we put all the pieces together and sell it to our cloud partner, who manages the deployment for you - getting you the asset before anyone else can get their hands on it.
What profile of customers have you delivered for so far? What has the impact been?
We've delivered capacity to hyperscalers, Fortune 500s, leading inference providers, and even NVIDIA themselves. We've got a really broad capability to pinpoint the right places to put compute, get the right deals on servers, and get the prime operator in place for a successful deployment and scaling plan.
The impact is obviously less burden on their teams to solve this operational headache themselves, but it’s also ensuring competitive edge.
Compute is a strategic resource that fundamentally enables or inhibits a company’s ability to scale and improve their AI product. Anthropic and Cursor’s recent compute deals with SpaceX show how the leaders in this industry are thinking about this. That’s strategic positioning to ensure they have the resources they need to train the best models and deliver them to the most users. When we deliver resources for our customers that no one else can, we’re giving them that same advantage at a different scale.
In our last conversation, you described Shadeform as the "Kayak for compute" - transparent on providers and pricing, with a high-touch support layer. A year on, what's changed about how you differentiate? With Capacity Solutions in the mix and the market more crowded, where does Shadeform's edge sit today?
Candidly, a big part of our platform is the transparency we've brought to the market and the trust we've built with partners. We still provide transparency on provider availability. We have a single provisioning solution for managing infrastructure across 30+ clouds. We also offer managed Kubernetes, managed Slurm, and storage solutions.
Shadeform Capacity Solutions strengthens our ability to deliver at scale for larger organizations and helps smaller teams get ahead of crunches that might impede their growth. Like I said, we've gained the trust of organizations from leading neoclouds to hyperscalers to NVIDIA, and that's a big part of our credibility and moat.
Anyone can try to deliver infrastructure. The question is whether you can do it reliably, with trusted partners, at scale, across geographies, in a distributed fashion with 30+ clouds and suppliers. There's a lot that goes into creating that unified experience across such a wide range of suppliers and clusters.
Shadeform unifies supply from some of the most well known neoclouds - Nebius, DigitalOcean, Lambda, Crusoe, and others. From the cloud provider side of the marketplace, why do they choose to work with Shadeform? What does Shadeform unlock for a GPU provider that they can't easily build themselves?
First and foremost, it's commercial. What we deliver to our cloud partners is demand from the top consumers in the industry - right now with pretty limitless capacity needs. We provide tremendous transparency around that. We often let our partners know the end customer, work with them to scale the end customer, and can even work on a referral basis when it's the right mix.
The second part is reputation. We've built a reliable reputation for delivering quality infrastructure and being transparent and ethical about our pricing and strategy. We don't dangle carrots in front of providers to get them to hold infrastructure for us speculatively. We're matter-of-fact about what we need and what we deliver. We don't play tricks, and if you survey our partners, that's something we believe they'd say resoundingly.
The Shadeform Update
In Feb 2025, you broke the user base into hobbyists, fast-growing startups, and enterprise teams. What's shifted in that mix over the last year? Are there segments that have leaned in harder than you expected, or new ones that have emerged?
The biggest shift has been the dramatic increase in enterprise and hyperscale consumption. Long-context and agentic workloads have pushed token needs up across just about every organization, and that's pulled in customers we weren't really serving a year ago.
Fortune 500s are the clearest example. Last time we spoke, we weren't working with many of them. Now that list includes top media companies, top banks, top technology MSPs and everyone in between. What's surprised us most is how quickly they've moved. These are organizations that traditionally take a long time to adopt new infrastructure, and they're now some of our most active customers.
The more surprising shift has been on the supply side of the market. Hyperscalers and neoclouds weren't really customers a year ago, but their own demand has grown so fast that they're now sourcing additional capacity through us to keep up.
AI-native inference providers, which were always a core segment, have continued to scale aggressively. We work with most of them at this point.
Before we wrap up, what are a few things you're most excited about over the next 6-12 months? Could be Shadeform-specific or where you think the broader compute market is headed.
On the Shadeform side, I'm excited about the continued evolution of accelerators and TCO efficiencies - getting cost-per-token down and improving the economics for our customers. Generally, that's where we want to provide most of our value: being the most affordable way to put together all the compute solutions needed to deliver.
On the market more broadly - what's always been eye-opening about this industry is how much projections get shattered. It's a moving target, and the market moves at its own whims. One thing I'm interested to see is how shifts in compute ownership and utilization at the largest scale reverberate down to AI-native companies and traditional businesses.
Stay up to date on the latest with Shadeform, get started with Shadeform Capacity Solutions, and connect with the team here.
Read our past few Deep Dives below:
If you would like us to ‘Deep Dive’ a founder, team or product launch, please reply to this email ([email protected]) or DM us on Twitter or LinkedIn.