Here's what's on our plate today:

• 🧪 Why AI money is quietly moving down the stack.

• 📰 An OpenAI drug-discovery spinout, Meta's AI layoff lawsuit, and New York's data-center freeze.

• 🛠️ Weekend To-Do: price out open-model inference, read the neocloud thesis, test a model router.

• 🗳️ Poll: Where does the lasting value in AI actually accrue?

Let’s dive in. No floaties needed.

The Laboratory

TL;DR

The models get the podium, but the tires get paid every lap.

• The $8.3B bet: Together AI raised $800M at an $8.3B valuation, led by Aramco Ventures, with NVIDIA participating, more than sixfold above its March 2024 valuation. It has never built a frontier model.

• Why investors care: Enterprises are cutting token spend, and the companies that make AI cheaper to run are becoming the most valuable.

• The catch: Anthropic still spends over half on at least one major routing platform. Frontier Labs’ own discovery, open source owns production.

• What's at stake: If model quality commoditizes, the durable asset stops being the weights and becomes the infrastructure they run on.

Why AI money is moving down the stack

Every Formula 1 season, 11 teams compete across 24 races to decide which one will be crowned world champion. Yet while the teams and the fans fixate on that one winner, Pirelli wins every race, not because it competes, but because, regardless of which car crosses the finish line first, it is running on its tires.

This kind of business is commonly known as a ‘picks-and-shovels’ infrastructure business, and a version of it is quietly emerging in the AI industry. In July, the market priced that business model when Together AI raised $800M at an $8.3B valuation. Much of the discussion focused on whether that valuation was justified. The more revealing story, however, is what the round says about where investors increasingly believe the enduring value in AI will be created.

A billion dollars of bookings without a model to sell

Together AI's latest funding round, led by Aramco Ventures with participation from NVIDIA, Vista Equity Partners, and General Catalyst, offers a window into where investors increasingly believe AI's long-term value will be created. The company belongs to a growing category known as neoclouds: specialized cloud providers built around GPU infrastructure, the graphics processors that power AI workloads. Its business is centered on inference, the process of running trained AI models to generate responses for real-world users and applications.

The company has never built a proprietary frontier model of its own, and that is precisely the point. Instead, Together AI provides the infrastructure that allows businesses to deploy and scale open-weight models such as DeepSeek and Kimi, whose parameters are publicly available for anyone to use and modify. According to the company, annualized bookings surpassed $1.15B last quarter, while open-weight model usage across the industry tripled over the past year. Those figures have not been independently audited, but they help illustrate why investors are increasingly willing to back companies that supply AI infrastructure rather than compete in the race to build the models themselves.

The more revealing signal, however, is how investors have repriced the business itself. Together AI was valued at $ 1.25B in March 2024 and $3.3B in February 2025, before reaching $8.3B in this latest round, indicating the market has increased the value of the same underlying idea by more than sixfold in roughly 28 months.

The trajectory and rate of this growth matter because Together AI is not the only company investors have valued this way. In the weeks before Together AI’s funding rounds, Upscale AI raised $500M at a $2B valuation, while TensorWave secured $350M at a $1.55B valuation. Both these companies operate in the same way as Together AI, which runs AI rather than building its own frontier models. Looked at individually, these are funding announcements. Looked at together, they begin to resemble a market that repeatedly expresses the same view.

The signal underneath the round

A funding round is ultimately a bet. Unlike a press release, which reflects what a company wants the market to believe, a funding round reveals what investors are willing to commit capital to. The more interesting question, then, is what those investors believe they are actually buying.

To answer this question, one needs to examine the timing of the announcements. Five days before Together AI announced its funding, CNBC reported that enterprise customers of OpenAI and Anthropic were shifting away from unconstrained token spending toward cost efficiency. One startup CEO described his decision to move all traffic from Claude to DeepSeek as a matter of survival. The same report quoted analyst Gil Luria as saying, "Current growth rates for Anthropic and OpenAI are the fastest they will ever be," a remark that was as much a warning about slowing growth as an observation about arithmetic, particularly given that both companies are preparing for eventual public listings.

Taken together, the two developments point toward the same conclusion. As enterprises have become more disciplined about the cost of running AI, the companies helping them lower those costs have become some of the most valuable businesses in the ecosystem.

In this scenario, companies like Together AI are betting that inference, not the models themselves, will become the enduring source of value. The company says customers can reduce inference costs by between sixfold and sixtyfold compared with relying on proprietary models, while CEO Vipul Ved Prakash describes its mission as ensuring that "intelligence is abundant, not expensive." Abundance is an unusual promise for a technology company because abundant goods rarely command premium prices, yet that is precisely the wager investors appear to be making. If intelligence itself becomes increasingly commoditized, the enduring business may belong not to whoever creates it, but to whoever can deliver it at the lowest cost, much as pipeline operators ultimately outlasted many of the wildcatters whose oil they carried.

The tension the thesis has to survive

The thesis is compelling, but it also has to explain one stubborn fact. In the very same week Together AI announced its funding round, TechCrunch examined token-spending data across major AI routing platforms and found that Anthropic still accounted for more than half of enterprise AI spending on at least one of them, even as open-source models had overtaken proprietary alternatives in raw usage volume. One investor quoted in the report offered perhaps the cleanest description of the market now taking shape, arguing that frontier labs would continue to own discovery while "open source will increasingly own production."

That observation deserves unpacking because it reframes the story from one of replacement to one of specialization. In this reading, the AI economy is splitting into two layers rather than tipping decisively from proprietary models to open ones.

Frontier laboratories remain where companies discover new capabilities and solve the hardest problems, while neocloud providers increasingly handle proven workloads that need to run cheaply and at scale. The former earns premium pricing for cutting-edge capabilities; the latter competes on efficiency.

McKinsey's latest survey points toward exactly this balance: more than half of organizations already use open-source AI somewhere in their stack, and 76% expect that usage to increase, even as they continue paying for proprietary models where governance, reliability, or cutting-edge capability justifies the premium.

However, despite investors' faith and its increasing importance in the AI space, the business of neocloud providers is not without its problems.

What could break the machine?

The neocloud model has its own fragilities, and they reinforce one another in ways that make the thesis less straightforward than the funding headlines suggest. Running inference at scale consumes capital almost as aggressively as building frontier models, and Together AI plans to expand its infrastructure footprint roughly fiftyfold over the next five years. Achieving that scale will require substantially more than the $800M it has just raised, with future investment effectively betting on demand in a market that, in anything resembling its current form, has existed for barely three years.

None of those caveats invalidates the investment thesis. They simply remind us that investors are not buying certainty; they are buying a particular vision of how the AI market will evolve.

The race that the money thinks it is watching

Return to the Formula 1 grid for a moment. Pirelli's business works because the sport eventually settled into a structure where cars change, champions rotate, and tires remain. The investors behind Together AI are betting that AI is moving toward a remarkably similar equilibrium in which models become the rotating champions. In contrast, inference infrastructure becomes the permanent tires underneath them. That wager could still prove wrong in either direction, because a decisive breakthrough from a frontier laboratory could restore proprietary models as the industry's scarce asset, while a genuine collapse in model differentiation could compress inference margins along with everything else.

What the company's $8.3B valuation ultimately tells us is that serious money believes neither outcome is the most likely, and instead bets on a market with many competing models, no permanent champion, and one class of companies that benefits regardless of which model wins. Whether those models ultimately prove to be the champions of the AI era or simply the vehicles running on someone else's infrastructure is the question the next several years of capital allocation will answer, and the honest entry in the lab notebook is that nobody, including the investors who just wrote the checks, knows the answer yet.

Headlines You Actually Need

• OpenAI's $2B drug-discovery spinout: OpenAI researcher Miles Wang is in talks to leave and launch an AI drug-discovery startup, reportedly raising around $200M at a $2B valuation with Lightspeed set to lead.

• Meta's AI layoff lawsuit: Twenty-six former Meta employees are suing in California federal court, alleging the company's AI-powered ranking tools disproportionately flagged workers with disabilities or medical leave for its mass layoffs.

• New York freezes data centers: Governor Kathy Hochul signed the nation's first statewide moratorium on new hyperscale data centers, pausing permits for up to a year while the state writes rules to protect the grid, water, and ratepayers.

Friday Poll

Weekend To-Do

• Price out open-model inference: Run one task you currently run on a proprietary API on an open-weight model via a neocloud like Together AI to feel the cost gap yourself.

• Read the neocloud thesis: Skim the recent CNBC and TechCrunch reporting on enterprises shifting token spend toward efficiency, the clearest picture of why inference is suddenly being repriced.

• Test a model router: Try a routing tool that sends each request to the cheapest, capable model, using the same efficiency logic that pushes the money down the stack.

Meme Of The Day

The Toolkit

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

• Tabnine: AI coding assistant that runs privately on your stack, useful for teams that can't send code to public AI services.

• Dust: No-code platform for building custom AI agents that connect to your company's tools and data, so teams can automate workflows without engineering.

Rate This Edition