An 8-Person Team Chases 1 Million AI Cloud Users
Zo Computer wants to give consumers something they've never had: their own private, always-on cloud computer. The company builds a personal AI cloud where users get servers, databases, and files managed by an always-on agent—a full computing environment, not just a chatbot. People use it to run small businesses, track health data, organize finances, and more, often without realizing the complexity underneath.
"Cloud is one of the best computing models of all time, and consumers have zero direct access because it's so complicated," said Rob Cheung, co-founder and CEO of Zo. "Now, with AI, it's finally possible for all of us to have cloud computers."
Zo is a two-and-a-half-year-old, 8-person company with an ambitious 2026 target: onboard one million new users. That goal means handling millions of model calls daily, and users who text their agent like a friend expect instant responses.
"We're building a new model of personal cloud computing that's always-on, elastic, and private by default for every user. Vercel gives us the AI infrastructure to make it possible."
Rob Cheung co-founder and CEO @ Zo Computer
The Adapter Code Burden
Zo offers access to any model users want, including bring-your-own-key support. That means their backend must talk to every major provider—OpenAI, Anthropic, MiniMax, GLM, Fireworks, and others. Before switching to Vercel, this required custom adapter code per provider, each with different image handling, key management, and edge case quirks.
Zo's own engineers also manually managed retries, provider routing, and fallback logic. Every new model release meant writing a new adapter, testing edge cases, and running a full deployment. With models launching weekly, this constant maintenance was a drag on a small team building a consumer product.
The impact showed up in their numbers. Zo's AI model calls ran at a 98% success rate with a 7.5% retry rate, meaning 1 in 50 messages failed or retried—tens of thousands of model fallbacks every day.
"We didn't even know what we were missing until after we switched to Vercel's AI Gateway. The revelation came through the numbers. We just had so many failures previously."
Rob Cheung co-founder and CEO @ Zo Computer
Two Layers, One Integration
Zo adopted both the AI SDK and AI Gateway, each solving a separate issue. The AI SDK replaced their custom adapter code entirely, offering a unified interface for every model—from image support to response format normalization—without per-provider implementations. The AI Gateway took over infrastructure complexity: retries, fallback routing, provider health monitoring, and uptime handled at the routing layer instead of inside Zo's codebase.
Ben Guo, Zo's co-founder, previously built APIs at Stripe where developer experience was the product. He sees the combined effect of both layers the same way: everything just works, and the hidden pieces matter most.
"Moving to the gateway is just so ergonomic. We get references to model names, and then rely on you to do the correct implementations and handle the edge cases."
Ben Guo co-founder @ Zo Computer
Supporting a new model went from an hour-long, multi-file code change to adding a config string in 30 seconds. When MiniMax released M2.7, Zo had it live the same day—no adapter, no edge case testing, no deploy cycle.
Measuring the Reliability Gain
Zo ran Vercel and non-Vercel routes simultaneously during rollout, creating a live A/B comparison under identical production conditions.
The results were stark:
Period | Route | POST error | Chat success | Retry rate | Avg attempts |
|---|---|---|---|---|---|
Before switch | Non-Vercel | 4.59% | 99.73% | 7.52% | 1.12 |
After switch | Non-Vercel | 10.38% | 97.86% | 17.07% | 1.29 |
After switch | Vercel | 0.45% | 99.93% | 0.34% | 1.00 |
While the non-Vercel route degraded over the test period, the Vercel path held steady. Retry rate fell from 7.5% to 0.34%—a 20x improvement—and average attempts per chat hit 1.00, meaning nearly every request succeeded on the first try.
For MiniMax M2.5, Zo's most-used model, latency improved significantly in an apples-to-apples comparison over the same window. Vercel handled 18,139 chats versus 21,105 on the non-Vercel route and still outperformed it:
Average latency improved 25.7%
P95 improved from 46s to 34s (25% improvement)
P99 improved from 131s to 81s (38% improvement)
The P99 number matters most because Zo's users text their agents throughout the day. A 131-second worst-case wait breaks that kind of interaction entirely.
"131 seconds to wait for something is just terrible. Now we can get 99% of our requests in under 80, which is huge."
Rob Cheung co-founder and CEO @ Zo Computer
By the end of the test, 91.88% of Zo's traffic routed through Vercel, which handled 3.3x larger context windows (42,500 average input tokens vs. 12,700) at a lower error rate than the non-Vercel path.
Clearing the Path to Growth
With AI SDK and AI Gateway handling the AI layer and Vercel hosting their marketing site, Zo no longer maintains adapter code. The team of eight can focus on their product rather than plumbing.
Keeping pace with rapid AI model development was once a source of worry for Cheung.
"We're a tiny team, so we want to spend our effort in the right ways. It's really nice to lean on Vercel and trust that we can add hundreds of times more traffic."
Rob Cheung co-founder and CEO @ Zo Computer
