The Race We’re Actually Running, Part 2: Good Enough at a Fraction of the Price

extract

CERN Partners With Mistral to Build a New Kind of AI for Physics CERN, the world’s largest particle physics laboratory, has spent decades collecting enormous amounts of data... What the AI race is a contest over, and why the cheapest capable model may win In part 1 of this series, I looked at where the US and China stand after two decades of investment. This installment asks the question the public debate tends to skip: when people say “winning the AI race,” what does winning mean? The AI race gets described as though there were one finish line: one model, one moment when someone crosses first and the outcome is settled. The United States and China are not running toward the same point. If China is racing toward nine o’clock and the US toward two o’clock, both can run as fast as they possibly can and still not compete for the same prize. The useful question is what each side is trying to get, and what happens to the rest of the world once they get it. Frontier model capability. This is the headline contest: which model scores highest on the hardest benchmarks. The US leads, narrowly. Stanford’s 2026 AI Index put the gap between the leading US and Chinese models on the Arena leaderboard at about 2.7% as of March. In mid-October, the Artificial Analysis Intelligence Index scored the best closed model at 58 and the best openly downloadable model at 46. The pattern is consistent: close on conversation, wider on hard reasoning and long autonomous tasks. Chip manufacturing. The US holds a wide and, for now, structural lead: advanced chip fabrication sits in a small number of facilities the US and its allies control. Part of that lead is policy. The Institute for Progress estimates that with no H200 exports, the US would hold a 21 to 49 times advantage in AI compute produced in 2026, and that unrestricted exports of Nvidia‘s H200 could cut it to roughly 1.2 to 6.7 times. Commerce approved about 10 Chinese buyers in May, but Reuters reported no deliveries as Beijing steered its companies toward Huawei. That substitution is not one for one. Huawei’s Ascend 910C trails the H200 in computing power and memory bandwidth, and Epoch AI estimates Huawei’s 2026 output at about 880,000 H100-equivalents against about 23 million for Nvidia, under 4%. The tightest constraint is high-bandwidth memory, which China cannot yet make in volume. Huawei is pushing anyway. In September it pulled its Ascend 960DT forward to early 2027, a chip Bernstein estimates will roughly match the H200 with more than twice its interconnect bandwidth, and said a 256,000-card cluster is being deployed. How large the US lead stays depends on Washington’s export decisions and China’s memory production as much as on chip design. Research volume and open-source distribution. China leads here, and this may be the contest that decides what the rest of the world runs on. Chinese models accounted for about 41% of downloads on Hugging Face over the past year, ahead of US models, and the summer brought a run of Chinese releases: Moonshot‘s Kimi K3, Alibaba‘s Qwen3.8, DeepSeek V4.1-Flash, and models from Z.ai and Xiaomi. In October, Anthropic reported that Z.ai’s open-weight GLM-5.3 came close to its restricted Mythos Preview model on cyber-exploit tests, while NIST’s Center for AI Standards and Innovation put it about four months behind the US frontier. DeepSeek and Huawei have also released open tools for programming Huawei’s chips, aimed at reducing dependence on Nvidia’s CUDA software. A note on terms. An open-weight model publishes its trained parameters, called weights, so anyone can download it, run it on their own hardware, and modify it. Many people call these open-source models, though the training data and code usually stay private. A closed model, such as those from OpenAI, Anthropic, and Google, can only be used through the company’s own service. The trade between the two is simple. Closed models are more capable, and open models are much cheaper. DeepSeek’s open-weight V4-Pro costs $0.66 per million input tokens and $1.98 per million output tokens off-peak (a token is about three-quarters of a word), against $5 and $25 for Anthropic’s Claude Opus 4.7 and $5 and $30 for OpenAI’s GPT-5.5. That makes DeepSeek roughly 8 to 15 times cheaper. On SWE-bench Verified, a widely used test of fixing real software bugs, Opus 4.7 solves 87.6% of the problems; DeepSeek reports 80.6%, and an independent US government evaluation measured 74%. Put simply, the open model solves fewer of the hardest problems for about a tenth of the price. Open models also reach what the best closed models could do only recently. Epoch AI puts the lag at about four months. So a buyer can pay a premium for the top of the range now, or give up some capability on the hardest tasks at a fraction of the cost and expect open models to reach today’s frontier within months. Many will take that trade for customer service, document processing, and routine coding, where the cheaper model already does the job. Open models are not only a Chinese story, either: hosting companies offer OpenAI’s open-weight gpt-oss-120b for well under a dollar per million tokens. Capital. Stanford’s AI Index puts US AI investment in 2025 at about $286 billion against about $12 billion for China, roughly 23 times. China answers with state capital. Bloomberg reported in June that Beijing has drafted a five-year, roughly $295 billion plan for a national AI data-center grid, funded by state bonds, requiring at least 80%domestic technology and interconnected by 2028. That is about $59 billion a year, less than any one of the largest US cloud companies spends: Oracle, Microsoft, Amazon, Meta, and Alphabet together are expected to spend about $800 billion on capital projects in 2026, according to research by Columbia Business School’s Stijn Van Nieuwerburgh. The Chinese plan’s strength is coordination, bringing state debt, land, power allocation, and domestic chips together under one national grid. Talent. This trend favors China and compounds over the longest horizon, which makes it the hardest to reverse quickly. The most visible diplomacy of the fall was the Trump-Xi summit in Washington, which ended September 25. According to news reports of the readouts, it extended the trade truce to January 10, 2027, eased tariffs on about $30 billion of non-sensitive goods in each direction, and opened a dialogue on AI risks with a channel for AI-related incidents. Rare-earth talks will continue, and advanced chips did not feature. The leaders are expected to meet again at APEC in November and the G20 in Miami in December. The diplomacy is moving on its own track, and the competition in chips, models, and power is moving faster. If you get to nine o’clock and I get to two o’clock, and we both got there as fast as we could, what did either of us win? People who study this closely suggest there may be no single winner in the sense the phrase implies. The more plausible outcome is an asymmetric split. The US likely keeps its lead in frontier capability and chip manufacturing, while China leads in research volume, open-source distribution, and the compounding effect of a much larger talent base. Asymmetric does not mean inconsequential. One framing among analysts is that if China wins, it will look like procurement and not like conquest. It will look like a government IT office in Jakarta, Lagos, or São Paulo choosing the system that is good enough, priced aggressively, translated into the local language, and bundled with financing and support, the way much of the world ended up on Chinese telecom infrastructure over the past decade. Infrastructure adoption rarely announces itself. It accumulates one procurement decision at a time until the alternative is gone. The stake, then, is a slow decision made country by country and company by company over the next decade about whose infrastructure the world’s economy runs on. Infrastructure carries the values, incentives, and leverage of whoever built it. A government or company that depends on another nation’s AI stack has handed that nation influence that is very hard to take back. That split may not hold as cleanly as it appears. On September 3, Nvidia announced it will acquire Hugging Face, the platform where developers find, share, and run open-source AI models, for $12.93 billion, its second-largest deal after a roughly $20 billion asset and licensing agreement with Groq in December 2025. Hugging Face reportedly brings in about $150 million a year, so the price reflects strategic position more than earnings. Nvidia expects to close in the first half of 2027, pending regulatory approval, and Jensen Huang has said Hugging Face will remain open to the entire AI ecosystem, with no requirement to use Nvidia compute. If it closes, it complicates the argument that China leads distribution. A US company at the center of open-source hosting, alongside the best available compute, would put American industry on both ends of the market: the frontier tier, and the distribution rails the inexpensive, everywhere tier runs on. Regulators will weigh chips and distribution under one roof. Nvidia’s last deal of comparable size, the $40 billion Arm acquisition, was abandoned in 2022 after regulators in several countries challenged it. Nvidia’s own filing lists the risk that regulators could bar Chinese models from the site, which could push Chinese developers toward domestic platforms such as Alibaba’s ModelScope and split open-source distribution in two. An earlier era of high-performance computing offers a useful precedent, and we lived it. Enterprise IT and supercomputing treat a processor in opposite ways. In a corporate data center, one hardened HPE or Dell server may carry hundreds of jobs, so it is built for reliability: high availability, N+1 redundancy, disaster recovery. Once it is stable, the team posts guards around it and touches it only for security upgrades, hoping nothing breaks. In a supercomputer, one processor may be one of a thousand working on a single job. In the CPU-cluster era, we built those machines from racks of inexpensive commodity servers, nicknamed pizza boxes for their thin, stackable shape, in place of a few exquisitely engineered machines from HP, Dell, Cray, or SGI. Each node was a worker. With proper checkpointing, if one got tired and sat down, the others picked up its work, and if one won the lottery and left for good, we sent it to the graveyard without much regret. It had already earned its keep by getting us to answers faster. That allowed a different strategy: build the most capable supercomputer possible for the money. A lower price for a set amount of performance meant we could scale larger and reach answers sooner on the same budget. Once a system was running and producing answers, we started again immediately, hunting the next bottleneck in the network, cache, memory, compute, or parallel storage. Find it, break the system, fix it, get more performance for the price, then break it again, without stopping. At Gulfstream, that loop showed up as faster and more accurate engineering answers and quicker production, in a business where the goal is to fly higher and faster than anyone else. The GPU and AI-chip world may be heading toward the same logic, which is why price-performance may end up a more decisive metric than peak benchmark performance. Huang said in May that GPUs Nvidia sold four or five years ago are now rising in price, and in August he described A100s, first shipped in 2020, as mission-capable through 2029. If durability and secondary-market value make older, cheaper compute economical for a large share of workloads, the same logic applies with more force to compute that was cheaper from day one. This is also where the hallucination question resolves. A model does not need to match the best frontier model on its hardest tasks to be good enough. Its error rate needs to fall below what the specific workflow can tolerate. For a first-pass literature review, customer-service routing, or routine coding, many non-frontier models clear that bar at a fraction of the cost. For a cancer-treatment protocol or a nuclear engineering calculation, the bar stays far higher, and narrowly scoped domain-specific AI has the advantage over any general-purpose model. An AI trained only on vetted, authoritative nuclear engineering literature will hallucinate far less on nuclear questions than one trained on the entire internet, because the good source material is all it has to draw from. In naval nuclear power, an operator does not act on a single gauge he cannot cross-check against a second, independent indication. AI used for high-stakes work needs the same discipline: knowing where every answer came from and how much that source can be trusted. The contest between general frontier capability and narrow, high-integrity domain capability will likely shape the next phase more than any single benchmark. It points toward specialized, verifiable agents built on clean, authoritative data and coordinated by higher-level agents, each layer improving the reliability of the one below it. That compounding is why neither country can afford to fall far behind. Once a system built this way starts improving its own tools and training data, the gap between leader and follower grows. Part 3 covers the fronts running just behind the frontier-model race, robotics and physical AI and cybersecurity, and the engineering answers to the water, noise, and land concerns driving opposition to data centers. About the Author: Larry Patterson works where high-performance computing, AI, and energy meet. A former Navy reactor operator aboard USS Enterprise (CVN-65), he founded and led Gulfstream Aerospace’s Scientific Research and Innovation SuperComputing Center from 2005 to 2020, where power, cooling, and cloud strategy were daily decisions. He has worked with the nuclear division of Continuum Dynamics. He is founder and CEO of HEAT Holdings, a Service-Disabled Veteran-Owned Small Business developing patent-pending ways to finance nuclear power and AI compute together, with every source of data behind each deal rated for quality. Related Item: The Race We’re Actually Running, Part 1: Twenty Years, Traded Places Editor’s note: The opinions expressed in this article are those of the author and do not represent the views of HPCwire.