The Cadence Trade
Why Anthropic and OpenAI are about to pull away
In early June, Wharton's Ethan Mollick published a short observation about AI labs. It read like a passing thought, but it was the first clear articulation of something that's been visible in the data for six months.
If AI self-improvement, even in a very limited way, is possible, the cadence of shipping both AI products, harnesses, and models should go up. This appears to be happening at Anthropic and OpenAI, but not for any other labs, including those that seemed to be catching up last year. Ethan Mollick, June 19, 2026
What Mollick was describing is a recursive loop. Models help build the next models. Coding agents engineer the training infrastructure for their own successors. The release cycle isn't just fast, it's getting faster because each release improves the tools that produce the next one.
This is what offline recursive self-improvement looks like at the lab level. The agent isn't the model. The agent is the lab. The clock is the release cadence. And starting in early 2026, two specific labs crossed a threshold the others haven't.
Cumulative count of major frontier model releases per lab from Q1 2023 to Q2 2026. Slope is cadence. Source: public release timelines, lab announcements.
The slopes are the story. Anthropic and OpenAI lines are steepest, and their slopes are bending upward. Google sat flat through most of 2025 and only began catching up in Q2 2026, when it finally started shipping multiple releases per quarter. Meta plateaued after Llama 4 in April 2025 and has shipped zero frontier models in the fourteen months since. DeepSeek runs a consistent quarterly cadence from a 2024 start but doesn't accelerate.
The diverging slopes are what makes this a recursive signal rather than a shipping race. Anthropic and OpenAI aren't just shipping faster. They're shipping faster at an accelerating rate.
Annualized release rate, computed as a rolling four-quarter window. A flat horizontal line means linear cadence. An upward-bending line means recursive cadence. Two labs are bending.
The second chart makes the acceleration concrete. Anthropic shipped two to three frontier models per year through most of 2024. By Q2 2026, that rate had doubled to six per year. OpenAI's rate climbed from two to five over the same period. Google held steady at two per year through 2025 and only crossed four in Q2 2026, after the cohort had already started pulling away. Meta's rate is in decline.
A linear improver shows a flat horizontal line on this chart. A recursive improver shows a line that bends upward. Two labs are bending. Three aren't.
Why this is RSI, not just shipping
There's a flat version of this story where Anthropic and OpenAI are just better-funded or better-managed, and the cadence advantage doesn't compound. The argument for why it does compound rests on three independent observations.
The first is that the labs use their own products to build their successors. Anthropic engineers use Claude Code to write Claude's training infrastructure. OpenAI uses Codex to ship Codex's next version. Each release tightens the tooling that produces the next release. The harness improves between releases, so the next release ships faster and better. That's recursion through the deployment layer, not the weights layer. The model itself stays frozen between versions. But the lab as a whole, the system that produces the model, is updating itself continuously.
The second is talent gravity. In June 2026, two of the most significant moves of the decade landed in the same week. Noam Shazeer, the co-author of the original Transformer paper and the architect behind multi-query attention and modern MoE routing, joined OpenAI as lead for architecture research. John Jumper, the AlphaFold scientist who shared the 2024 Nobel Prize, left Google DeepMind for Anthropic after nine years. Zachary Lipton left DeepMind the same week for an unspecified destination. The single best architecture researcher and the most decorated scientist in modern AI both moved to labs already running the tightest cadence loops. Talent follows shipping velocity, and shipping velocity benefits from talent. That's compounding.
The third is compute. Tri Dao's FlashAttention-4, released in March 2026, hit 71 percent utilization on NVIDIA B200 hardware. That's a generational improvement in attention efficiency, and it's tuned for the inference workloads that dominate post-training and agent rollouts. Mamba-3, also released in March 2026, was the first major architecture release explicitly designed inference-first rather than training-first. Both signal the same thing: each next generation of models gets cheaper to train and cheaper to run. Lower compute cost per training cycle means more training cycles per quarter, which feeds back into faster cadence.
When cadence, talent, and compute compound at the same labs simultaneously, that's the parabolic setup.
What goes parabolic, exactly
Anthropic and OpenAI are still private. So the question of what goes parabolic routes through their compute substrate, their cloud partners, and their tooling moats.
| Beneficiary | Why |
|---|---|
| NVDA aggregate demand | Faster cadence equals more training runs plus more inference rollouts. Both lab compute curves bend up. |
| NVDA share of AI capex | The fast labs spend disproportionately at NVDA. Slow labs spread across vendors. |
| MSFT | Public proxy for the OpenAI loop. GitHub is the dev substrate, Azure is the compute layer. |
| Anthropic valuation | Already swyx-floated at $2T target. Conservative if cadence holds. |
| Power basket | CEG, VST, GEV. Compounding inference needs compounding electricity. |
The first-order beneficiary is NVIDIA. Faster cadence means more training runs and more inference rollouts. Both lab compute curves bend up. The two labs running the tightest loops also spend the most concentrated dollars at NVIDIA. Slow labs spread their capex across vendors. Fast labs converge.
The second-order beneficiary is Microsoft. MSFT is the public proxy for OpenAI's RSI loop. GitHub is the development substrate where the loop closes. Azure is the compute infrastructure. The Cursor Origin Git announcement is a real threat to GitHub's developer mindshare, but the OpenAI loop runs on GitHub for now, and the longer it does, the harder it is to displace.
The third-order beneficiary is power. Constellation, Vistra, GE Vernova, the whole AI-aligned power complex. Compounding inference needs compounding electricity. Power is the slowest-moving piece of the buildout and the hardest to replace.
Anthropic's IPO, when it happens, will price somewhere between the swyx-floated two trillion dollar target and whatever the cadence math implies if the loop holds. That number is conservative if the slope keeps bending.
What kills the case
Three things could break this trade.
Continual learning ships first at a lagging lab. If Google, DeepMind, or an open-source effort ships true online recursive self-improvement before the offline-RSI labs do, the moat inverts. The slow labs would jump straight to the next paradigm and the cadence advantage of the loop-runners stops mattering.
A genuine architecture paradigm shift lands outside the cohort. If pure-Sutton agents, scaled Mamba-3 systems, or whatever comes after current transformers ships from Tri Dao's Together AI, Albert Gu's Cartesia, or an academic lab, the incumbents lose architectural advantage even if they keep shipping at the current cadence.
The loop breaks for regulatory reasons. The week of June 19, 2026, Anthropic released Fable 5 with built-in restrictions that prevented developers from using it to build competing LLM technology. The U.S. Commerce Department then issued export controls on Mythos and Fable, requiring licenses for foreign nationals. Andrew Ng wrote a sharp open letter arguing that the whole stack of restrictions, including Anthropic's own developer restrictions and the government's export controls, was forcing other nations toward AI sovereignty alternatives. If safety incidents or regulation interrupt deployment, the loop stops compounding.
These are real risks. They aren't the base case. But they're the variables to watch.
The window
The chart isn't proof that AI is improving. The chart is proof that two specific labs have crossed the threshold where their own products materially shorten their own development cycle, and three others haven't. That's the recursive self-improvement threshold. Once crossed, the slope keeps steepening until something external stops it.
The window where this is visible but not yet priced is now. Once the next two quarters print and the slope keeps bending, it's consensus. The trade is sized to the gap between visible to you and consensus.
The single cleanest expression is long NVDA plus the power basket, with MSFT as a corollary on the OpenAI loop. The single cleanest fade is the labs that aren't in the cohort, which means Google equity until the cadence visibly closes, Meta until Llama 5 ships, and any data-licensor business whose moat depends on static information rather than environment-coupled learning.
The Mollick observation isn't a prediction. It's a description of a pattern that already started.