Discovery vs. Production: The Two-Tier AI Economy Taking Shape

For two years, the debate over open-source AI has been framed as a zero-sum contest. Either the frontier labs hold their moat, or cheap open-weight models eat their market. Every DeepSeek release triggers the same question: is this the moment Anthropic and OpenAI start bleeding?

A new framing published this week suggests the question itself is wrong, and the usage data backs it up.

On Monday, Decagon CEO Jesse Zhang published a post titled "Everyone is wrong about open source AI in the enterprise." His observation starts from an apparent contradiction inside his own company. Mature AI deployments at Decagon are switching to lighter, cheaper models. Yet overall spend on expensive frontier models has barely moved.

Zhang's resolution: frontier and open-source models are not competitors. They are two phases of the same life cycle. Expensive frontier models prove out new use cases. Once a use case matures and stabilizes, it migrates to cheaper open-source alternatives for production at scale. Meanwhile, new use cases keep arriving at the frontier. As Zhang puts it, "The frontier labs will keep owning discovery. Open source will increasingly own production."

What the Data Shows

The picture in the routing data is striking, and it splits cleanly along the volume-versus-value line.

On Vercel's AI gateway, DeepSeek has surged to the lead in token volume, processing just over a third of all tokens on the platform, with Chinese lab Z.ai's GLM-5.2 climbing to fourth. But scroll from volume to spend and the picture inverts: Anthropic still accounts for more than half of total AI spend on the platform.

OpenRouter, which captures a larger and somewhat less enterprise-heavy slice of the market, tells the same story. DeepSeek V4 Flash processes 5.3 trillion tokens weekly, more than double the roughly 2 trillion handled by the most popular frontier model, Claude Opus 4.8. But Opus is priced at roughly 23 times more per token, $1.37 per million versus 6 cents. On simple arithmetic, the frontier model still captures the lion's share of the spending despite handling a fraction of the volume.

Open source is winning tokens. Frontier is keeping the money. That is not a market being disrupted. That is a market stratifying.

The Capital Is Already Pricing This In

What makes Zhang's framing more than a clever blog post is that the investment flows of the past six weeks map onto it almost perfectly.

On July 1, Together AI, the leading cloud for running open-weight models like DeepSeek, Nemotron, Kimi, and GLM, closed an $800 million Series C at an $8.3 billion valuation, led by Aramco Ventures with participation from NVIDIA, Salesforce Ventures, and others. Its annual bookings crossed $1.15 billion, and open-weight usage on its platform tripled over twelve months. The rest of the inference-cloud field is riding the same wave: Groq raised $650 million in June, RunPod hit a $1 billion valuation, Baseten was valued at up to $13 billion, and Fireworks AI is reportedly in talks at $15 billion, nearly four times its October valuation.

Here is the detail that ties it together: Decagon, whose CEO wrote the post, is a named Together AI customer and reported cutting its inference costs sixfold after moving mature workloads to open models. The production layer of the two-tier economy is not theoretical. It is a billion-dollar bookings business growing fast enough to attract sovereign capital.

And yet none of that capital is flowing away from the frontier. It is funding a second, parallel layer of infrastructure underneath it.

Why the Frontier Holds

Two explanations, not mutually exclusive, account for the frontier labs' resilience.

The first is market growth. The universe of AI-addressable tasks is expanding so quickly that frontier labs can maintain revenue simply by dominating early-stage deployments. Every enterprise experimenting with a new agentic workflow, a new coding assistant, a new compliance process starts at the frontier, because that is where the capability ceiling is. Discovery is a permanently renewing market as long as capability keeps advancing.

The second is task difficulty. Some workloads never migrate down. The hardest problems, complex multi-step agentic work, novel reasoning, high-stakes domains, remain beyond what lighter models reliably handle. This maps directly onto the enterprise data we covered earlier this year: OpenAI's enterprise revenue crossing 40% on the back of "teams of agents," and Anthropic's enterprise traction, which as we argued in our governance brief is partly a story about trust, auditability, and reliability rather than raw benchmarks. The premium tier is not just smarter. It is more accountable, and regulated buyers pay for that.

There is also a subtler dynamic worth naming. The "discovery" phase is where pricing power lives. TechCrunch's Russell Brandom noted that as recently as last September the prevailing worry was that foundation labs would end up "selling coffee beans to Starbucks," commodity inputs to an application layer that captured the value. Part of that came true, since vertical AI plays did shift to lighter models. But token for token, the frontier providers have held onto the most desirable segment of the market: the premium-priced, hardest, newest workloads.

The Strategic Questions This Raises

The two-tier framing is elegant, but it leaves open questions that will determine how stable this equilibrium actually is.

How long does a use case stay at the frontier? The economics of the frontier labs depend on the discovery phase being long and sticky. If open models close the capability gap faster, the migration window shortens, and with it the period of premium monetization. NVIDIA's Nemotron, poised to leap up the leaderboards on the strength of NVIDIA's distribution and the model's adaptability, is exactly the kind of entrant that compresses that window. So is the sixtyfold cost reduction some Together AI customers report versus closed-model APIs.

Does discovery alone support frontier economics? Frontier labs are carrying extraordinary capital costs. A business model where customers systematically graduate away from you once their workloads mature is viable only if new discovery demand keeps arriving at a rate that offsets the churn. So far it has. That is an empirical question, revisited quarterly, not a law of nature.

Where does the enterprise governance layer sit? Our view, consistent with the thesis we laid out in the agentic finance governance brief, is that the migration from frontier to open source is gated by more than capability. Regulated deployments need identity-linked agent authorization, audit trails, and explainable decisions. Open models can match capability long before the surrounding governance tooling matches what enterprises need. That friction extends the frontier's hold on regulated workloads specifically, which happen to be among the highest-value workloads in the market.

What This Means for Markets

Three observations for institutional allocators.

First, stop treating open-source model releases as a bearish signal for frontier labs by default. The volume data and the spend data are measuring different markets. A DeepSeek surge in token share tells you the production layer is scaling. It tells you very little about frontier revenue, which is concentrated in discovery-phase and hard-workload spend. The two-tier structure means the right question for any AI exposure is which tier a company monetizes, not how many tokens it serves.

Second, the production layer is now independently investable, and heavily capitalized. Together AI, Groq, Fireworks, Baseten, and RunPod collectively represent a new infrastructure category: the inference clouds that serve open-weight models at scale. Their economics resemble cloud computing more than model labs, competing on cost, reliability, and serving efficiency. The $800 million Aramco-led round suggests sovereign and strategic capital views this layer as durable infrastructure, not a residual of the frontier race.

Third, watch the migration rate as the key variable. The stability of the two-tier economy rests on the pace at which use cases graduate from frontier to open source versus the pace at which new frontier use cases emerge. If capability advances slow while open models keep improving, the discovery tier shrinks and the "coffee beans to Starbucks" scenario returns. If agentic and regulated workloads keep expanding, the frontier premium persists. That ratio, more than any single benchmark, is the number to track through the rest of 2026.

Zhang's framing will not settle the open-source debate, and the "yet" in TechCrunch's headline is doing honest work. But it replaces a zero-sum question with a structural one, and the structure it describes matches both the routing data and the capital flows. Discovery and production are becoming different businesses, with different economics, different competitors, and different winners. Allocators should underwrite them separately.