Here’s what’s on our plate today:

• 🧪 Why AI aficionados should be looking closely at Poolside.

• 📰 AI race in the wrong place, Karpathy joins Anthropic, Google's I/O charm offensive.

• 🛠️ Three tools worth trying: Laguna XS.2, Ollama, and Modal.

• 🗳️ Poll: What does Poolside's collapse tell us about AI's future?

Let’s dive in. No floaties needed…

The Laboratory

TL;DR

• Models outran the grid: Poolside, a $3B coding-AI startup, planned a 2 GW Texas data center with CoreWeave, but the deal collapsed under CoreWeave’s financial strain and Poolside’s unproven track record.

• The model shipped anyway: Weeks later, Poolside released Laguna XS.2, an open-weight coding model that beat GPT-4o on SWE-bench, verified with far fewer resources.

• Power is now a filter: Nearly half of planned U.S. data center builds for 2026 are delayed or canceled. Hyperscalers are spending $725B to keep up; startups can’t compete for electricity.

• The real stakes: If efficiency gains keep shrinking compute needs, startups like Poolside rent their way to relevance. If frontier training still demands massive scale, AI consolidates around the few companies wealthy enough to build their own power plants.

Why AI aficionados should be looking closely at Poolside

For most of the past three years, the AI conversation has centered on models: who trained the biggest one, which lab posted the highest benchmark score, and whether the next generation would be meaningfully smarter than the last. While that era is not over, it has acquired a companion problem that increasingly determines what gets built and by whom. The problem in question has shifted from being purely algorithmic to increasingly physical with electricity, land, cooling water, grid interconnection queues that stretch four to five years, and transformers with lead times measured in years rather than months.

Which means the AI industry’s ambitions have run headlong into the material limits of the world, and the collision is producing a new kind of company story, one where a startup’s most revealing moment is not its model release but the data center that never got built. Poolside is that story in concentrated form.

The company and the plan

Poolside is a foundation model company founded in 2023 by Jason Warner, formerly GitHub’s CTO, and Eiso Kant. Headquartered between Paris and New York, it builds AI models trained specifically for software engineering, the thesis being that a model built from scratch on code, rather than adapted from a general-purpose language model, can outperform the incumbents at the tasks developers actually do. In October 2024, the company raised $500M in a Series B at a $3B valuation, backed by Bain Capital Ventures, Bezos Expeditions, and NVIDIA.

However, what came next was more unusual. Rather than renting compute from Amazon, Google, or Microsoft, Poolside built its own infrastructure from the ground up. In October 2025, the company announced Project Horizon, a planned 2 GW data center campus on 568 acres in Fort Stockton, Texas, with CoreWeave as anchor tenant. The first phase was to deliver 250 MW of capacity under a 15-year lease, scaling to 500 MW, with CoreWeave supplying more than 40k NVIDIA GB300 NVL72 GPUs. A parallel $2B Series C, with NVIDIA reportedly anchoring up to $1B, was meant to finance the build. Poolside’s co-founder and co-CEO Eiso Kant framed the ambition in the CoreWeave announcement: “To compete at the frontier, you need to be vertically integrated from dirt to intelligence.”

How the deal fell apart

By late 2025, the partnership had begun to unravel as Poolside was unable to stand up its first cluster of chips on CoreWeave’s timeline. CoreWeave, meanwhile, was navigating its own pressures. The company had gone public in March 2025 at $40 per share, pricing below its indicated range of $47 to $55. While the stock eventually climbed past $180 by mid-year, it then fell more than 60% by December amid concerns about its debt load, which exceeded $14B, and construction delays at its own facilities.

CoreWeave was not in a position to take long-term, single-tenant bets on a startup that had not yet shipped a publicly benchmarked model. And the company ultimately chose to pursue different paths for its own strategic and timing reasons.

With the anchor tenant gone, the planned $2B Series C quickly unraveled. According to the Financial Times, investors were unconvinced that Poolside could realistically train models capable of competing with frontier labs like Anthropic, OpenAI, and Google DeepMind. Poolside later explored reviving a smaller 400 MW portion of the project in discussions with Google, but those talks reportedly lost momentum by mid-April 2026, leaving the 568-acre site still idle.

The model that shipped anyway

Less than a month after the infrastructure collapse became public, Poolside released its Laguna model family. The centerpiece was Laguna XS.2, an open-weight coding model released under the permissive Apache 2.0 license, allowing developers to use and modify it freely.

Built using a mixture-of-experts (MoE) architecture that activates only small specialized portions of the model for each task, Laguna was efficient enough to run locally on high-end consumer hardware while still posting competitive results on SWE-bench Verified, a benchmark that measures whether AI systems can solve real software engineering problems from GitHub repositories.

Poolside’s larger proprietary model performed even better, placing the company unexpectedly close to frontier systems developed by far larger rivals like OpenAI. According to VentureBeat, Poolside’s models outperformed GPT-4o on the same benchmark despite being developed with a fraction of the resources.

In other words, Poolside’s software capabilities appear to have advanced faster than the infrastructure needed to support them at scale, and that widening gap between what AI companies can build and what they can realistically power may be the more important story.

The tension this exposes

Poolside’s situation is unusually stark, but the underlying physics apply to everyone. The AI industry in 2026 is defined by a mismatch between how fast models improve and how slowly the physical infrastructure to train and run them can be built.

According to Fortune, U.S. data center development has hit what one Wood Mackenzie analyst called “a bend in the trajectory,” because utilities lack grid and generating capacity to build fast enough. Data centers accounted for roughly 50% of all U.S. electricity demand growth last year, according to the IEA, and that share is expected to hold through 2030.

The scale of capital being deployed is staggering but still insufficient. Google, Amazon, Microsoft, and Meta collectively plan to spend $725B on capital expenditures in 2026, up 77% from the prior year, according to Tom’s Hardware. And yet, nearly half of planned U.S. data center builds for 2026 have been delayed or canceled, largely due to infrastructure limitations. Microsoft, which committed roughly $80B to AI infrastructure this year, has reportedly begun weighing whether to delay its clean energy matching goals because the data center buildout is colliding with those targets. According to an SEC proxy filing by shareholder advocacy group As You Sow, Meta’s record-breaking ‘Hyperion’ data center in Louisiana has drawn significant community opposition because it requires power from seven new gas plants. And across 18 states, policymakers have passed or introduced moratoriums on new data center construction.

The difference between these companies and Poolside, however, is structural. The hyperscalers face the same physics, the same grid queues, the same community resistance, but they have balance sheets that can absorb years of delay. Microsoft can partner with Chevron to build a 5 GW natural gas plant in West Texas, but a startup with $500M in the bank cannot. For Poolside and companies like it, the infrastructure bottleneck is not just a delay: it is a filter that determines who gets to train frontier models at all.

A counterargument worth taking seriously

There is a credible argument that the data center collapse may ultimately clarify Poolside’s business rather than destroy it. By most outside assessments, the company’s attempt to become vertically integrated “from dirt to intelligence” was an expensive distraction for a startup that had not yet proven its models could compete. Now that Laguna has demonstrated credible performance, Poolside can focus on what it does best, model architecture and training, while renting compute the way most AI labs already do.

The open-weight XS.2 release also appears to be a deliberate ecosystem strategy similar to Meta’s approach with Llama and Mistral AI: use open models to attract developers and funnel adoption toward enterprise products. If the models are strong enough, computing becomes a financing problem rather than an existential one.

However, that argument is limited in scope because frontier training runs now cost tens to hundreds of millions of dollars, while rivals from GitHub, Google, Amazon, Cursor, and Windsurf are competing for the same customers. In that environment, a strong model released today must be followed by a better one within months, and that still requires computing that Poolside does not control.

What this says about the path AI is on

The broader picture becomes systemic the moment you step past Poolside. The American power grid was not built for the kind of concentrated, high-density load that AI data centers represent. According to Gartner, power shortages could restrict 40% of AI data centers by 2027. The effects are already reaching ordinary households: in Nevada, NV Energy told the utility serving nearly 49k Lake Tahoe residents that it will stop providing them power after May 2027, in part because it needs the capacity for data centers being built by Google, Apple, and Microsoft.

The hyperscalers are increasingly becoming energy companies themselves, financing power generation, building on-site energy systems, and negotiating directly with utilities. That adaptation is possible for companies capable of spending tens of billions of dollars. It is not available to startups like Poolside, whose model development depends on infrastructure they cannot meaningfully control.

AI models are becoming more capable and more efficient, with systems like Laguna XS.2 showing that strong coding performance can be achieved with relatively modest compute. But efficiency alone does not resolve the industry’s deeper constraint: training the next generation of models still requires electricity, chips, and infrastructure at a scale that startups, and increasingly even large corporations, cannot reliably secure.

Whether that asymmetry consolidates AI development into the hands of a few hyperscalers, or whether the efficiency gains demonstrated by models like Laguna make the mega-data-center approach unnecessary, is something no one in the industry has answered yet, and that Poolside, sitting on 568 empty acres in West Texas, is now living through in real time.

Quick Bits, No Fluff

• The AI race in the wrong place: A new SAP analysis argues the AI race is being fought in the wrong place, with too much focus on models and not enough on the workflows that actually create value.

• Karpathy joins Anthropic: Andrej Karpathy, OpenAI co-founder and one of AI's most influential researchers, is joining Anthropic's pre-training team, a major signal shift in the talent war.

• Google's I/O courts both sides: Google is expected to use its I/O conference to court both developers and consumers, balancing AI coding tools with mainstream Gemini features.

Thursday Poll

3 Things Worth Trying

• Laguna XS.2: Poolside's open-weight coding model runs locally on high-end consumer hardware and rivals GPT-4o on real software engineering benchmarks.

• Ollama: Easiest way to run open-source models like Laguna, Llama, and Mistral on your own machine, no cloud bills required.

• Modal: Serverless GPU platform for training and inference, useful if you want to test models at scale without committing to long-term cloud contracts.

The Toolkit

• Leonardo AI: AI image and video generator with fine-grained creative controls, built for designers, marketers, and game studios who need consistent style at scale.

• Modal: Serverless cloud for running Python and AI workloads, lets you spin up GPUs in seconds without touching infrastructure.

• Quillbot: AI writing assistant that paraphrases, summarizes, and rewrites text on demand, useful for tightening drafts or escaping your own voice.

Rate This Edition