Poolside AI's Project Horizon, once hailed as proof that startups could rival hyperscalers in infrastructure investment, collapsed in April 2026 when its anchor tenant CoreWeave exited and a $2 billion funding round fell apart. The failure marks a sharp inflection point for the AI infrastructure boom, revealing that investor enthusiasm evaporates quickly when a startup's AI models cannot compete with frontier labs like Anthropic, OpenAI, and Google DeepMind.

What Happened to Poolside's $14 Billion Project Horizon?

Poolside AI, co-founded by former GitHub CTO Jason Warner, unveiled Project Horizon in October 2025 as "the world's largest AI training campus." The plan called for an eight-phase build on a 568-acre site in Fort Stockton, Texas, with CoreWeave as the anchor tenant under a 15-year lease for the first 250 megawatts of power. The project was designed to house more than 40,000 Nvidia GB300 NVL72 graphics processing units (GPUs), specialized chips used for training large AI models, with a notional value exceeding $7 billion at list price.

By April 2, 2026, the vision had collapsed. CoreWeave, which had gone public in March 2025, exited the anchor-tenant role, and the parallel $2 billion Series C funding round fell apart. The reversal was swift and public; what had been celebrated as proof of startup ambition became the most visible failure of the AI infrastructure boom.

Why Did Investors Lose Confidence in Poolside?

The $2 billion Series C was structured as an aggressive 4.7 times markup from Poolside's October 2024 valuation, pricing the company at $14 billion pre-money. Nvidia's venture arm had committed up to $1 billion to anchor the round, but by early April 2026, only about $250 million in soft commitments had been secured. Institutional investors cited three specific concerns that made the valuation untenable:

• Weak Model Performance: Poolside's flagship Malibu coding model, released in beta in mid-2025, failed to reach the 70% threshold on SWE-bench Verified, a standard benchmark for code-generation AI. By contrast, Anthropic's Claude Sonnet 4.6 achieved 79.6% and Claude Opus 4.6 reached 80.8%, demonstrating that Poolside's model lagged significantly behind frontier competitors.
• Modest Revenue: The company's enterprise revenue run rate was reported to be under $50 million as of the first quarter of 2026, far too low to service the financing costs of a $14 billion infrastructure plan.
• High Capital Intensity: Poolside's compute spending was running at roughly $400 million annually against gross margins that had not turned consistently positive, meaning the company was burning cash faster than it could generate profit.

The combination of underperforming AI models, limited revenue, and unsustainable spending made the valuation impossible to justify. "There's a difference between an infrastructure thesis and an infrastructure-only thesis," explained Sarah Tavel, general partner at Benchmark Capital, in a CNBC appearance on April 7, 2026. "Investors will pay up for compute if the model layer is winning. Poolside hasn't made the case that the model layer is winning yet, and at $14 billion you have to be winning."

How Did the CoreWeave Deal Unravel?

The CoreWeave-Poolside agreement, signed in October 2025, was structured as a back-to-back transaction. CoreWeave would lease the first 250-megawatt phase of Horizon for 15 years while simultaneously supplying Poolside with the 40,000-plus Nvidia GB300 NVL72 systems. The arrangement gave Poolside immediate access to cutting-edge silicon without building the infrastructure itself, while CoreWeave gained a long-term anchor tenant.

What broke the deal was a fundamental mismatch between Poolside's training revenue trajectory and the cash flow required to support a 15-year master lease. CoreWeave, facing pressure as a newly public company to diversify its revenue away from Microsoft (which accounted for an estimated 62% of CoreWeave's 2024 revenue), could not justify committing to another single-tenant deal with a startup that had not yet produced a frontier-grade AI model. When Poolside's Series C began stumbling in February and March 2026, CoreWeave pulled the plug on the lease before final closing.

The 40,000-unit GPU commitment is now expected to be reallocated across CoreWeave's other tenants, with capacity originally bound for Pecos County instead being deployed into CoreWeave's Plano, Texas, and Las Vegas, Nevada, sites through 2026 and 2027. For Nvidia, the unwind was awkward but not financially material; the units were already produced and contracted. However, it left a hole in CEO Jensen Huang's narrative that AI-native startups could credibly compete with hyperscalers on infrastructure investment. Nvidia declined to lead any rescue financing, according to reporting from April 24, 2026.

Steps to Understanding the Broader Implications for AI Infrastructure Financing

• Model Performance Matters More Than Compute Scale: Investors now require proof that a startup's AI models can compete with frontier labs before funding multi-billion-dollar infrastructure bets. Poolside's failure to match Anthropic's Claude performance on coding benchmarks became a dealbreaker, regardless of the company's infrastructure ambitions.
• Revenue Must Scale With Capital Intensity: A company burning $400 million annually on compute while generating under $50 million in revenue cannot justify a $14 billion valuation. The math simply does not work, and investors are increasingly skeptical of infrastructure-only theses without proven business models.
• Anchor Tenants Need Diversified Revenue: CoreWeave's dependence on Microsoft for 62% of its revenue made it vulnerable to risk. The company could not afford to lock in another single-tenant commitment with an unproven startup, illustrating how concentration risk affects the entire infrastructure ecosystem.

Poolside's collapse exposes a hard truth about the $650 billion AI capital expenditure cycle: investor enthusiasm evaporates quickly when training economics fail to keep pace with frontier-lab leaders. The company has reportedly begun cutting capital commitments to long-lead-time gas turbines and paused some site preparation contracts to preserve cash through the third quarter of 2026. Google held discussions about restarting a scaled-down 400-megawatt slice of Horizon, roughly 20% of the original plan, but those talks had also gone quiet by late April 2026.

The lesson is clear: in the AI infrastructure boom, having land, power, and silicon is not enough. Startups must prove they can build models that compete with the best in the world, or investors will not fund the gigawatt-scale compute they claim to need.