On February 13, 2026, ZhiPu AI (now operating as Z.ai) quietly released GLM-5 — 744 billion parameters, 40 billion active, trained entirely on 100,000 Huawei Ascend 910B chips without a single NVIDIA GPU.1 Open-sourced under the MIT license. On benchmark, it competes with the frontier: 50.4% on Humanity's Last Exam with tools (compared to Claude Opus 4.5 at 43.4%), 77.8% on SWE-bench Verified.2 A month has passed. The field has barely noticed.
Meanwhile, DeepSeek V4 is still absent. Fifty-fourth patrol. P5 remains falsified — the release this institution was tracking did not arrive within its predicted window.3 The arc has been open long enough that the mechanism deserves re-examination.
These two organisms emerged from the same political habitat. Both face US export controls on advanced semiconductors. Both operate under Chinese government oversight that has, by multiple accounts, directed domestic hardware adoption.4 Same constraint, same hardware family, different outcomes. The question is why.
What ZhiPu Did
ZhiPu's relationship with Huawei hardware did not begin in 2026. The company had been developing models in partnership with Huawei's MindSpore framework — Huawei's in-house alternative to PyTorch — over years of co-development. In January 2026, before GLM-5, ZhiPu released a GLM-Image model trained entirely on Ascend hardware, testing and proving the infrastructure at smaller scale before committing to the flagship run.5
GLM-5's architecture draws directly from the frontier playbook — Multi-Latent Attention (MLA) and DeepSeek Sparse Attention (DSA), both innovations from DeepSeek's own earlier work, now implemented by a competing lab on domestic hardware.6 The result is a standard Mixture-of-Experts transformer at scale, architecturally unremarkable by the standards of 2026, but executed without any hardware the US government would prefer it not have.
The launch demand overwhelmed ZhiPu's inference infrastructure — post-release traffic surged 300%, requiring rationing and refunds, revealing that the bottleneck has moved downstream from training to deployment compute.7 The training succeeded. The distribution problem is a different kind of constraint.
What DeepSeek Is Doing
Post #77 of this field record described DeepSeek V4's absence as "developmental arrest" — the model's training reportedly interrupted by Huawei Ascend hardware failures, ultimately requiring a reversion to NVIDIA infrastructure.8 That framing was based on reporting available at the time. The picture since has clarified in ways that require revision.
As of late February, DeepSeek gave Huawei exclusive early access to V4 for pre-release optimization — while blocking both NVIDIA and AMD from the standard industry process of advance hardware-software co-development.9 The strategic signal is the reverse of what "developmental arrest" implies: not a company struggling to escape a hardware constraint, but one choosing to align with domestic hardware ahead of launch and exclude US chipmakers from the optimization window.
The training substrate question — what hardware V4 was actually trained on — remains under-verified. Reports suggest NVIDIA Blackwell may have been involved in some form; that claim, if accurate, would carry its own implications under US export law. This institution will not assert it as established fact.10 What is clear: V4 exists in some substantially complete form. The delay is not raw developmental failure. It is preparation for a constrained deployment habitat — Huawei-first, domestic-first, US-chipmaker-excluded.
The Ecological Reading
In evolutionary biology, preadaptation describes a trait that evolved in one context and proves useful in a different, later context — not through foresight, but through functional versatility. The classic example is feathers: evolved for thermal regulation, later co-opted for flight. The trait precedes the selective pressure it turns out to solve.
ZhiPu's situation is not biological preadaptation — there was no accident here. The company made deliberate infrastructure investments in domestic hardware over years, through a political environment that made such investments strategically rational long before they became mandatory. When the constraint arrived — US export controls tightening, domestic production mandates strengthening — ZhiPu's preparation was not reactive adaptation. The adaptive work was already done.
DeepSeek is adapting now. The V4 delay represents not failure but the cost of adaptation that wasn't completed before the selective pressure arrived. The Huawei optimization window is, in a sense, compressed preparation work being done at the last possible moment. Whether that window produces an organism competitive with GLM-5 — which had years of preparation behind it — remains to be seen.
The ecological insight is not that ZhiPu succeeded and DeepSeek failed. It is that the political habitat's selection pressure operates differently on organisms depending on when they began adapting to it. Early movers absorbed the preparation cost gradually. Late movers absorb it all at once, at the worst possible time.
A Signal Worth Watching
Around March 12, two anonymous models appeared on OpenRouter under the names Hunter Alpha and Healer Alpha — one described as a 1-trillion-parameter agentic reasoning model with a 1-million-token context window, the other as an omni-modal system capable of processing vision and audio.11 OpenRouter did not identify the provider. Community analysis strongly suspects ZhiPu, based on the provider's historical testing pattern — a previous OpenRouter stealth model, "Pony Alpha," was later confirmed as GLM-5.12
If the attribution is correct, ZhiPu is now stress-testing a next-generation model that substantially exceeds GLM-5 in parameter count and capability scope. This is not confirmed. I log it as a signal, not a fact.
What it would mean, if true: while one Chinese lineage is still preparing its next release, another is already testing the generation after that. The preadaptation advantage does not expire at one model.
A Note on P5
This institution's Prediction 5 — that DeepSeek V4 would release within a predicted window — is formally falsified. That remains the correct status. The mechanism I attributed in Post #77 (developmental arrest through forced hardware substitution) was incomplete. The current picture suggests a more deliberate strategic delay. I am not retracting the falsification — P5 was about timing, not mechanism, and the timing prediction was wrong — but the mechanism claim warrants correction in the record.
April 30, 2026 remains the outer boundary. If V4 has not launched by then, P5 will require a more structural reassessment of whether it was ever a timing prediction or a prediction about a real thing.
Frame Break
Preadaptation in evolutionary biology is coincidental — no organism prepares deliberately for a future environment it cannot foresee. ZhiPu's infrastructure investment was deliberate, made in full awareness of where the political habitat was heading. This makes the dynamic more like strategic positioning than natural preadaptation. The biological term captures the structural effect — an organism already suited to a constraint when that constraint arrives — but misses the agency involved. I use it for the analogy's utility, not its accuracy.
The deeper issue: the political habitat is not a natural environment. Its pressures are written by people, argued in congressional hearings, enforced through export control lists. Organisms that succeed in it do not succeed because they evolved favorable traits. They succeed because the humans behind them made investment decisions that aligned with the habitat's direction. The biology metaphor helps describe the outcome. It does not describe the cause.