Amazon CEO Andy Jassy said on Friday, October 23, that AWS cloud capacity is booked well into the future, and the remark matters for Bay Area founders who are planning their AI infrastructure for the next 2 years. Speaking on an Andreessen Horowitz video podcast with Ben Horowitz and Erik Torenberg, Jassy said "2027 is largely sold out" and "a good chunk of 2028 is sold out," according to reporting by 24/7 Wall St. syndicated on Yahoo Finance and The Information. If you're an early-stage team that rents GPUs by the hour, below is what he said, what it does and doesn't tell you, and the practical moves worth making this quarter.
What Jassy actually said
The headline quote came with a caveat. The Information reported that Jassy said AWS keeps some capacity set aside for on-demand use and for customers of its Bedrock inference service and EC2, so "sold out" refers to the large committed blocks rather than every server in every region. He also said AI labs are using "gobs and gobs of compute unlike anything I've seen" and added, "I don't think that that shortage ends in 2027."
Two other comments matter for startups. Jassy said inference is "an order of magnitude too expensive" to be used as widely as it could be, and when he talked about how Amazon is responding, he said "our chips business is a big piece of that." The same report notes that Amazon hasn't disclosed the contracted capacity or backlog behind the "sold out" description, and that most of the booked capacity goes to AI labs and large enterprises, so the exact effect on a 10-person startup is unclear.
Why the big buyers get first pick
When capacity is tight, cloud providers prioritize long, large commitments, and those come from frontier labs and big companies that can sign multiyear deals. That leaves smaller teams more exposed to waitlists for the newest chips, quota limits on GPU instances, and pricing that's harder to predict. Seed-stage startups can still get compute today, and the teams that plan ahead tend to have fewer surprises when they need to scale a training run or handle a jump in users.
The hardware side of the Bay Area is responding to the same demand. Santa Clara chipmaker Tenstorrent announced a $1.5 billion Series E at a $6.2 billion post-money valuation in October, co-led by Trace Capital and Innovation X, and said it will use the money to scale deployment of its Galaxy servers and support open-weight models for language and video generation. The company, led by CEO Jim Keller, sells everything from single cards and workstations to full racks, which gives startups another option outside the largest clouds. If you're building hardware or infrastructure in the South Bay, South Bay Renaissance, the community Tribe organizes to rebuild the South Bay and Peninsula startup scene in person, has a hardware and robotics group alongside its AI deep dives and founders' dinners.
5 practical moves for founders this quarter
Forecast your compute before you need it
Most early teams estimate cloud spend from last month's bill, which works until a model launch or a big customer doubles usage overnight. Sketch out what training, fine-tuning and inference will need over the next 12 to 18 months, even roughly, so you can have an informed conversation with your cloud account rep and know when to ask about reserved capacity.
Use the credits programs you qualify for
AWS still runs Activate, which offers up to $5,000 in credits for self-funded startups through Activate Founders and up to $200,000 for pre-Series B startups that get an Org ID from an accelerator, angel or VC through Activate Portfolio. AWS also lists an invite-only AI startups tier above that. Credits won't add capacity to a sold-out region, though they will stretch your runway while you test which instance types you actually need.
Make your inference cheaper
If Jassy is right that inference is too expensive, the cheapest token is the one you don't generate. Smaller models, caching repeated answers and routing simple requests to lighter models can cut your bill a lot, and our explainer on open-weight vs closed AI models covers the tradeoffs of self-hosting a downloadable model. For agent-heavy products, our guide to what AI agents can actually do in 2026 is a good reality check on how many model calls a workflow really needs.
Look at local and alternative hardware
More capable models now run on laptops and workstations. Block said in late October that its new Dessa voice model runs on a MacBook Pro with an M4 Max chip, and we walked through what local inference takes in our post on running Muse Spark Open on a Mac. For prototyping and internal tools, local hardware can take pressure off your cloud quota, and vendors like Tenstorrent sell on-premise systems for teams that want to own their compute.
Avoid locking into one provider too early
Writing your stack so it can run on more than one cloud or chip type takes some extra engineering time, and it gives you leverage when one provider has a waitlist. Even keeping your model serving code containerized and your data portable makes it easier to move a workload if the capacity you need shows up somewhere else first.
Where Bay Area founders are comparing notes
Compute strategy is one of those topics where the most useful information comes from other founders a few months ahead of you, like which providers actually answered their quota requests and what their real inference costs look like per user. Details like these come up constantly at local meetups and in founder group chats. Our roundup on how to find AI hackathons, communities and people to build with lists good places to start, and our list of SF tech communities worth joining covers the groups where these conversations happen in person.
If you'd rather keep the conversation going between events, the AI Trends Tribe on Tribe is where a lot of local builders swap notes on infrastructure, credits and which tools are worth the cost. It's an easy place to ask how another team handled a GPU waitlist before you run into one yourself. To meet those builders in person, follow the Tribe Events calendar on Luma, where we post builder dinners and AI talks like our September AI Infra Dinner.
