On January 27, 2025, Nvidia lost $590 billion of market value in a single trading session — the largest one-day loss for any company in the history of public markets — because a Chinese lab most of Wall Street had never heard of a week earlier had published, for free, the weights to a model that looked like it belonged in the same conversation as America's best. The popular telling of what happened next runs backward. It says American open-source culture (Meta's Llama, the old Silicon Valley instinct to give the tools away and monetize the platform) was already a gift to the world, and China simply took it. The filings, the timestamps, and the people who were in the room that week tell a cleaner and less flattering story: the gift-giving framing has the order of operations exactly backward. DeepSeek didn't accept an invitation. It sent one.
Part 2 of Hormuz and the Weighting Game. Part 1 followed a researcher's return ticket to China and asked what tacit knowledge travels with a person that no export-control regime can seize. This part follows the thing that travels even faster than a person (a checkpoint, published once and downloaded forever) and asks who actually wins the race that publishing starts. The vertical axis (chips, power, equity) belongs to The Stack, cross-linked, not re-argued. This is the horizontal axis: who holds the weights, and on whose electricity bill they get retrained.
/ 01The Shock, By the Numbers
Start with the timeline, because the sequence is the entire argument. On December 26–27, 2024, DeepSeek released V3, an open-weight base and chat model, claiming a $5.6 million cost for its final training run: a figure that describes one run, not the lab's total infrastructure spend, and one that SemiAnalysis would later estimate against roughly $1.6 billion in total server capex.1 That distinction (final run versus total build-out) is the actual mechanism of what happened three and a half weeks later, because it is precisely the distinction the market did not make.
| Event | Detail | Date |
|---|---|---|
| DeepSeek-V3 released | Open-weight base/chat model; claimed $5.6M final-run cost, contested (SemiAnalysis: ~$1.6B total capex) | Dec 26–27, 2024 |
| DeepSeek-R1 released | Reasoning model rivaling OpenAI's o1, MIT license, technical paper disclosing method — not a black-box breakthrough | Jan 20–21, 2025 |
| Nvidia's single-day loss | −17% share price, −$590B market cap — largest single-day loss for any company, ever | Jan 27, 2025 |
| OpenAI's first open weights since 2019 | gpt-oss (120B/20B, Apache 2.0); coverage explicit it was “driven by competitive pressure from models such as Llama and DeepSeek” | Aug 5, 2025 |
| PRC models' share of global AI workloads | Per the House Select Committee on the CCP's own framing, in its letters opening a joint investigation into US firms' use of Chinese models | ~30% by end of 2025 |
Two reactions from that week are worth quoting directly, because they're the cleanest evidence of how the shock actually registered among the people whose job is to price it. Marc Andreessen, on the day of the crash: “DeepSeek R1 is AI's Sputnik moment.”3 Ben Thompson, writing at Stratechery the same week, sharpened the analogy in the direction that matters for this series: “this is like Sputnik if the Russians had made it to space and then explained in great detail how they got there.”1 The open publication (the technical paper, the MIT license, the fact that anyone could read exactly how DeepSeek did it) is what turned a single competitive model into a commoditization event. A closed frontier model that merely matched GPT-4-class performance would have been alarming. An open one, with the method disclosed, was structurally destabilizing, because it did not just prove the capability existed. It handed every other lab, American and Chinese, a working recipe for how to get there cheaper.
/ 02Sequencing the Blame Correctly
Here is the causal correction this part exists to make, and it matters because the wrong sequence is the more comfortable one to tell. It is comfortable to say American open-source culture (Zuckerberg's stated bet that Meta should “own the ecosystem instead of the access control”4) was already running before DeepSeek showed up, and that China's labs simply rode the wave America built. Llama's download numbers support part of that story: over a billion downloads by March 2025, 1.2 billion by April, 65,000-plus community fine-tunes.4 But the sequence of what happened in the seven months after the DeepSeek shock is not a story about an American gift being accepted. It is a story about an American retreat being forced.
The shock (Jan 20–27, 2025)
DeepSeek-R1 ships open-weight with a disclosed method. Nvidia loses $590B in a day on the theory that frontier-comparable performance no longer requires frontier-scale compute spend.
The admission (early Feb 2025)
Sam Altman, on Reddit: “I personally think we have been on the wrong side of history here and need to figure out a different open-source strategy.” OpenAI's own CEO concedes the closed-weight posture is losing, in public, within two weeks of the shock.5
The war rooms (same week)
Zuckerberg convenes four internal engineering “war rooms” to reverse-engineer DeepSeek's efficiency gains: two on training-cost technique, two on data composition. Meta's own infrastructure director reportedly warns colleagues DeepSeek could outperform not the current Llama, but the next one.6
The policy response (July 2025)
America's AI Action Plan reverses the prior administration's open-model caution and directs the government to actively promote American open-weight models as a geopolitical tool: explicitly aiming to make US models “the gold standard… worldwide” so that “allies are building on American technology.”7
The concession, made flesh (Aug 5, 2025)
OpenAI ships gpt-oss, its first open weights since GPT-2 in 2019. Six years of closed-weight strategy ended in six months, with contemporaneous coverage naming the exact competitive pressure that forced it.1
The retreat (Apr 2026)
Meta, the standard-bearer of American open weights, reverses course entirely: its Superintelligence Labs ships Muse Spark, Meta's first proprietary closed-weight flagship, breaking from the Llama lineage. The reported reason: Llama 4 Maverick had been overtaken on benchmarks by Chinese open models. The lab that founded the American open-weight lane exited it because China now sets the pace inside it.19
Read that sequence in order and the ideology-as-origin story doesn't survive it. An administration does not write a national AI strategy in reaction to its own country's prior generosity. A CEO does not concede publicly, on Reddit, that his company was “on the wrong side of history” because his own strategy was working too well. Every American open-weight move that followed January 27, 2025 (the Action Plan, gpt-oss, the doubled-down Llama roadmap) reads correctly as a participant's response to a race someone else started, not as the race's opening move. DeepSeek didn't inherit the American open-source ethic. It ambushed a market that thought it still controlled the pace of its own commoditization, and the pace question is the one that actually decides who benefits.
/ 03Who Absorbs Whom
Once the race is correctly sequenced, the next question is who wins it. And “wins” here does not mean whose model tops a given month's leaderboard. It means who converts an absorbed capability, of any origin, into a shipped product fastest and cheapest. On that question the record is asymmetric, but not in the simple “China copies, America invents” shape the original framing risked. It is asymmetric in velocity.
The clearest single case: in November 2024, Reuters reported that six researchers across three Chinese institutions (two under the PLA's Academy of Military Science) had built “ChatBIT,” a military intelligence and decision-support chatbot, on top of Meta's Llama 2 13B. Meta confirmed the use was unauthorized and called the underlying model “outdated.” That's a fair defense of the specific version, and also beside the actual point: an open American model, released under a permissive license Meta could not enforce at the point of use, became infrastructure for Chinese military AI with no mechanism for Meta to detect or prevent it.9 Running the same dynamic in the other technical direction: on January 30, 2025, OpenAI told reporters it had evidence DeepSeek-linked accounts used “obfuscated third-party routers” to query its models at scale and harvest outputs for training, a direct violation of terms of service that explicitly bar training a competing model on OpenAI's outputs. OpenAI's own language accused DeepSeek's effort of “free-riding on the capabilities developed by OpenAI and other US frontier labs”10, a genuinely notable irony from a company that itself trained on a scale of scraped web data it never fully disclosed, now the aggrieved party alleging IP theft in reverse. Both directions of absorption are real. Neither is symmetric in scale. And the extraction channel did not stay a one-week news story: a year later both major US labs escalated it into quantified federal complaints. In February 2026, OpenAI formalized its accusation in a memo to the House Select Committee on the CCP, and Anthropic filed its own technical disclosure naming DeepSeek, Moonshot AI, and MiniMax, counting roughly 24,000 fraudulent accounts and more than 16 million exchanges used to query Claude for apparent distillation purposes. Part 4 carries the full treatment of what those harvested exchanges contain; what matters here is the scale, which is no longer alleged in the abstract but counted.20
| Flow | Figure |
|---|---|
| Qwen-based derivative models, Hugging Face | 151,448 — 2.6× Meta-derived models, 4.7× Llama-family specifically |
| Chinese-origin models' share of top new-model downloads, 2025 | >45% (Qwen + DeepSeek + GLM + Kimi combined) |
| Chinese model inference pricing (MiniMax, Zhipu) | ~$0.30 per million input tokens, vs. ~$5 for Claude Opus — roughly 16.7× |
| US commercial adopters of Chinese open weights, named | Microsoft (Azure/GitHub, R1), Amazon (Bedrock, R1), Cursor/Anysphere (Kimi), Airbnb (Qwen — CEO's stated reason: “fast and cheap”) |
The congressional response is itself evidence of how fast and how deep this went. In spring 2026, the House Select Committee on the CCP and the House Homeland Security Committee opened a joint investigation into Airbnb and Anysphere specifically over their use of DeepSeek, Alibaba, Moonshot AI, and MiniMax models; the committees' own framing, not an outside estimate, put PRC models' share of global AI workloads at roughly 30% by the end of 2025.8 Weights that leak out of the United States mostly travel through unenforceable licenses and alleged terms-of-service violations: leakage the originating lab cannot fully see or stop. Weights that flow into the United States travel through completely legal, fully disclosed commercial adoption, because they are released under more permissive terms and, increasingly, because they are simply cheaper to run. Neither path required anyone to steal anything. The second path is just faster, and it is faster because of what section 06 gets to.
/ 04The Honest Steelman, and Where It Runs Out
A fair accounting of America's open-source strategy has to grant its architects a real argument, made in good faith, by people with no reason to be wrong about their own logic. David Sacks, the sitting White House AI and crypto czar, has made the ecosystem-lock-in case as plainly as anyone: “A meaningful segment of the global market will prefer the cost, customizability, and control that open source offers. We want the U.S. to win this category too”; and, reaching for the analogy that has anchored Silicon Valley's self-understanding for three decades, “Permissionless innovation is how America won the internet.”13 America's AI Action Plan codifies exactly this logic as official policy, not just industry rhetoric: give the world American open weights, and the world builds its habits, its tooling, and its standards on an American substrate, regardless of which country's model tops a given month's benchmark. It is not a naive argument. It mirrors how Android and Linux won platform share by losing the license fee, and there is a real, defensible case that Llama's early lead in downloads and derivatives proves the logic wasn't fantasy: American open models really did seed the current global open-weight norm, and the distribution infrastructure underneath nearly all of it (Hugging Face, the safetensors format, the transformers library) is still, quietly, American-built plumbing that Chinese labs ship their own weights through.
Take Sacks's argument on its own terms and it is genuinely strong: mindshare compounds even when a rival tops this month's leaderboard, and a developer who has built five projects on Llama's tooling conventions doesn't necessarily migrate to Qwen's the day a new benchmark drops. If the payoff to openness is measured in multi-year ecosystem control rather than this quarter's win, the strategy could still be correct even while losing on points right now. That distinction (strategy versus current scoreline) is the honest form of the steelman, and it deserves to be stated before it's set aside.
Here is where the honest version of the steelman has to confront its own author. In July 2026, after Moonshot AI's Kimi K3 (a 2.8-trillion-parameter open-weight model) took the top spot on the Frontend Code Arena and scored at or near the frontier on multiple other benchmarks, Sacks himself called the result “concerning,” warned that defensive American policy reactions (data-center permitting bans, pre-approval requirements for model releases) meant the US was “tying itself in knots,” and said plainly: “this is how you lose the AI race.”14 This is not a critic's assessment. It is the architect of the ecosystem-lock-in steelman, eighteen months after making the case, conceding on the record that the strategy is currently losing on the merits, while still defending permissionless innovation as the right approach, just one being executed too cautiously and too late. The steelman survives as a claim about what the right strategy was. It does not survive as a claim about what is happening in 2026. Reported download and adoption numbers (Qwen ahead of Llama on Hugging Face, Chinese models above 45% of top new-model downloads, roughly 30% of global AI workload share) back Sacks's own July assessment more than they back his February one.
/ 05A Move, Not a Present
It helps here to name what an open-weight release actually is, in the vocabulary international-relations scholars use for exactly this kind of act, rather than treat it as either charity or theft. Thomas Schelling's account of strategic commitment supplies the sharper frame. An actor becomes more credible, not less, when it deliberately gives up the option to reverse course: burning a bridge, tying its own hands, manufacturing an irreversibility that a purely discretionary actor could never signal.15 Publishing a model's weights is precisely this kind of move. Once a checkpoint is public, it cannot be unpublished; no future policy shift, boardroom decision, or export-control tightening can claw it back. DeepSeek's R1 release, and every American open-weight release that followed it, functions less like a gift with strings attached and more like a commitment device: each lab signaling, to rivals and to its own government simultaneously, that it is in this race and cannot credibly threaten to leave it. That single-shot irreversibility sits inside a much longer game, though, and that's where Robert Axelrod's work on the iterated Prisoner's Dilemma becomes the more useful lens for the race as a whole. Axelrod's central finding (that cooperation becomes rational in a repeated game because of the “shadow of the future,” the durable expectation of retaliation or reward) only holds if that shadow is long and credible on both sides.16 China's five-year-plan continuity gives its labs and regulators a planning horizon that doesn't reset with an election. American AI policy, by contrast, has visibly reversed at each administration change over the past decade, which shortens the credible time horizon any single US actor can commit to, and, by Axelrod's own logic, should make the US side structurally less able to sustain a patient, iterated strategy and more prone to reactive, single-cycle moves. That is not an accusation of bad faith. It is a structural reading of why the same open-weight instinct that might have paid off over twenty years of steady execution instead produced a scramble.
This is also the honest answer to the more comfortable rival hypothesis: that the whole US-China AI relationship is really a symmetric prisoner's dilemma, a mutual hostage exchange dressed up in Sputnik headlines. The frameworks above don't refute that reading; they sharpen it into a truer, more specific one. Both governments are behaving rationally inside a shared, iterated game, and both feel the exchange: that much of the symmetric framing is correct. What the game-theoretic literature on this exact case (Zeng's 2025 documentation of the ChatGPT-to-DeepSeek action-reaction spiral is the clearest empirical version of it17) actually supports is not mutual advantage but mutual reaction on unequal footing: a real, mirrored security-dilemma spiral sitting on top of an asymmetric ability to convert what gets absorbed into deployed advantage. Feeling the exchange and benefiting from it are not the same thing, and the distillation numbers in section 03 are the tally for which side currently does more of the second.
/ 06The Meter Underneath
One question this part can only gesture at, because it belongs properly to Part 3: why does absorption convert to advantage faster on one side of the Pacific than the other? The short answer is electricity. American household power runs roughly $0.18 per kilowatt-hour against a Chinese household rate near $0.08. Chinese industrial rates, especially in the western provinces where much of the country's newest compute is sited near wind and solar, run lower still.18 Distillation and fine-tuning (the retraining step that turns an absorbed model, of any origin, into a shippable competitive product) is a compute-intensive, power-hungry process in its own right. Cheap, reliable domestic power doesn't just subsidize the original training run. It subsidizes the entire downstream loop of absorbing, retraining, and redeploying whatever capability shows up on either side of the export-control line, which is the actual mechanism connecting this part's argument to the energy axis Part 3 covers in full.
“The chip embargo had removed the hardware. But the hardware had already been mapped by a mind that no customs officer could detain.”
/ 07The Race, Correctly Timestamped
Put the sequence back together. DeepSeek-R1 shipped open-weight, with its method disclosed, on January 20, 2025, and by January 27 it had erased more market value in a single day than any company loss in history. Every American open-weight move that followed (Altman's public reversal, Meta's war rooms, the AI Action Plan, gpt-oss) is dated after that week, and each one is documented, in its own contemporaneous coverage, as a response to competitive pressure rather than an independent act of generosity. The steelman for America's open-source strategy is real and was argued honestly here, by its own architect, before it was set against his own later assessment that the strategy is currently losing. And the frameworks that describe why (Schelling's commitment device for what a published checkpoint actually is, Axelrod's shadow of the future for why an electorally discontinuous strategy struggles against a continuously planned one) land on the same conclusion the raw distillation numbers do: openness didn't originate in Menlo Park and travel outward as a gift. It originated in Hangzhou as a shock, and the race it forced runs on whoever's electricity bill is smaller. Part 3 follows the tokens themselves — where American inference demand actually routes, and why.