AI Agenda Live Recap: The Compute Squeeze and the Startup Playbook for Scaling

After a brief reprieve, artificial intelligence enterprises once again find themselves in a bidding war for graphics processing units. For frontier labs and hyperscalers, spending billions on training and inference is business as usual. But for startups and independent developers, compute costs can quickly swallow the bottom line.

At a recent panel hosted by The Information, Co-Executive Editor Amir Efrati spoke with two experts on how smaller players can scale profitably.

• Roman Chernin, co-founder and chief business officer, Nebius
• Rita Kozlov, vice president, product management, Cloudflare

What’s Driving GPU Demand?

Not since the “Hopper crunch” of 2022 has competition for GPUs been this intense.

“People were ready to go very hard just to get access to compute,” recalled Chernin. “We saw startups sign these crazy three- and four-year deals, not knowing if they would even still exist in that time frame.”

While Nvidia eased supply constraints, the shortage has returned—this time driven not by training foundation models but by the proliferation of products built on top of them.

“Foundational models have unlocked the creativity and possibility for nonprofessional developers to build digital assets,” said Chernin. “We’re now at this really exciting stage where we can finally see those products.”

The Startup Growth Playbook

For newer AI apps, the early stage often means leaning on closed-source frontier models to test product-market fit. But once traction comes, Chernin warned, the economics shift fast.

“When they get to that growth stage, they need to do a lot of work to not lose all their money on compute,” he said. That means optimizing model costs, latency and data usage, and routing intelligently.

At scale, most startups look for partners. Cloudflare’s global edge network reduces latency for user-facing tasks, while Nebius provides backbone compute for heavier workloads. “Most of these companies aren’t infrastructure companies,” Chernin noted. “They want us to help lower the barrier so they can execute at scale.”

Kozlov added that many teams are cutting costs by tailoring models. “You don’t always need the entire dataset of what a frontier model is trained on,” she said. “Choosing the right model for the right task is where a lot of optimization happens.”

Case Study: Higgsfield

Nothing tests a startup’s resilience like sudden virality. Higgsfield, a video-generation app, went viral on TikTok and X, attracting 11 million users only five months after their initial launch.

“They’re not OpenAI. They don’t have the brand recognition where people will wait,” said Chernin.

They also don’t have the budget to throw at GPUs. Instead, Higgsfield built a prediction algorithm to pre-provision compute without overprovisioning. They also manage demand by prioritizing higher-revenue regions and paying users with faster, higher-quality generations, while offering limited free access elsewhere.

“These are normal production application problems,” Chernin said. “It’s not just prompting. You build pipelines, optimize, and predict demand.”

When the Chips Are Up

The arrival of Nvidia’s Blackwell GPUs has sparked questions about whether the new hardware will upend the market. Chernin stressed that each chip generation expands possibilities rather than replacing the last.

“Ampere is still in use today, because some use cases don’t need more performance or memory,” he said. Blackwell, however, unlocks new frontiers in large-scale reasoning and image generation.

“Every generation lets us do something we couldn’t do before with reasonable performance.”

The Future of Enterprise AI

While consumer-facing AI apps have captured headlines, enterprise adoption has lagged. Kozlov argued that is beginning to change.

“In the past six months to a year, I’ve seen enterprises move a lot faster,” she said. Standards like Model Context Protocol and tool calling are making integration easier, while demand from both executives and grassroots developers is creating momentum.

Still, challenges remain. Security and compliance loom large, with chief information officers and chief information security officers wary of exposing sensitive data. Cloudflare, Kozlov noted, is focusing on sandboxed, zero-trust environments where employees can safely build their own AI-powered applications.

The risk-benefit calculation, however, is shifting.

“The risk of not playing becomes existential,” Kozlov said. “We’re seeing companies realize that sitting on the sidelines is no longer an option.”