NextFin News - China's frontier AI models now match American systems on practical software tasks at a fraction of the cost, and the performance gap has narrowed to single digits faster than Washington's export controls assumed. On the same day, London's Green Party proposed a 38% windfall tax on the biggest UK banks' profits to fund small-business relief. Together, the two stories frame the central tension of this moment in global finance: the economics of artificial intelligence are being rewritten faster than the institutions meant to tax and regulate them can keep up.
US equities shrugged off the geopolitical implications and leaned into the earnings cycle instead. The S&P 500 climbed 0.55% to 6,927.93 and the Nasdaq Composite rose 0.79% to 23,044.10, as investors rotated back into technology and AI names ahead of Nvidia's earnings report. Software stocks rallied after Anthropic rolled out "Claude Cowork" integrations with Google Drive and Microsoft Excel, easing fears that AI would replace rather than enhance existing productivity tools. The market's read was clear: the AI spending cycle is intact, even if the leader of that cycle is no longer unchallenged.
China's AI Catch-Up Is Real, and It Is Cheap
The evidence sits in the benchmark data. On the Terminal-Bench 2.1 leaderboard, which tests models on practical software tasks such as fixing bugs, setting up servers and managing files, OpenAI's models lead at 89.5%, with Anthropic's Claude Opus 5 at 89.1% and GPT-5.6 at 88.0%. Moonshot's Chinese model Kimi K3 scores 85.0%, ahead of Claude Fable 5 at 84.6% and Grok 4.5 at 81.6%. Alibaba's Qwen 3.8 Max sits at 81.3%. The data is as of August 12, 2026, from Artificial Analysis.
The cost gap is starker than the performance gap. DeepSeek's V4-Flash model costs about $0.14 per million input tokens and $0.28 per million output tokens, roughly 100 times cheaper than Anthropic's Claude Fable 5 on identical benchmark tests, according to Artificial Analysis. GPT-5.6 Sol averaged $1.86 per benchmark test, while Claude Fable 5 came in at $3.15 per test. Alibaba's Qwen3.8-Max, a 2.4 trillion-parameter system, is priced at $2.00 per million input tokens and $6.00 per million output tokens.
On the broader Intelligence Index, which combines coding, reasoning and workplace-task performance across nine categories, DeepSeek's V4-Flash scored 50 out of 100, tying Google's Gemini 3.6 Flash but trailing Moonshot's Kimi K3 at 57. The data is as of August 14, 2026.
The gap is not new, but its texture has changed. Stanford's 2026 AI Index report found that "the U.S.-China AI model performance gap has effectively closed," noting that US and Chinese models have traded the lead multiple times since early 2025 and that, as of March 2026, Anthropic's top model led by just 2.7%. The same report found that China leads the world in AI publication volume, citations, patent output and industrial robot installations, even as the US still produces more top-tier models and higher-impact patents.
How did Chinese labs get here under US chip restrictions? They stopped trying to win the hardware war on American terms. DeepSeek, Alibaba and Moonshot focused on systems that perform almost on par with leading US models without needing the most powerful hardware, betting heavily on open-weight AI software where internal parameters are made available for developers to share, study and tweak. That challenges the dominant US business model, which is predicated on controlling the frontier through proprietary scale and closed model access.
There is a darker channel in the same story. OpenAI and Anthropic have accused Chinese companies, including DeepSeek, of training new releases on the outputs of US frontier models through a technique known as distillation, using fraudulent accounts and proxy services to conduct what the US firms call coordinated "distillation attacks." Distillation can be a legitimate training method; the allegation is that it is being used at scale to extract capability from models that Chinese labs cannot legally access. If true, the catch-up is partly borrowed rather than homegrown — and if Washington cracks down on the proxy channels, the measured gap could reopen quickly.
The Hardware Constraint Has Not Disappeared
The counter-argument is serious, and it comes from the Council on Foreign Relations: Huawei is not a viable competitor to Nvidia, the best US AI chips are about five times more powerful than Huawei's best offerings, and the gap is projected to widen to seventeen times by 2027. CFR argues that US export controls should remain in place because they are working, and that loosening them — for example, exporting three million H200 chips to China in 2026 — would hand China more AI computing power than it could produce domestically until 2028 or 2029.
But the market is already routing around the constraint. Chinese AI firms have been accessing advanced Nvidia computing power through data centers in Southeast Asia, exploiting a loophole in the US export-control regime that focuses on the ownership of physical chips rather than remote access to computing power. Meanwhile, GLM-5, a 744 billion-parameter model, was trained entirely on Huawei Ascend chips, and DeepSeek V4 is optimized for Ascend. Huawei planned around 750,000 Ascend 950PR units in 2026, with mass production beginning in April and full-scale shipments expected in the second half of the year.
This is the mechanism: export controls raise the cost of Chinese AI, but they do not stop it. They shift training to domestic silicon where possible, to third-country data centers where necessary, and to efficiency gains everywhere. The result is a Chinese AI sector that is more expensive to build than it would otherwise be, but still advancing. China's 15th Five-Year Plan carries an aggressive AI infrastructure strategy built around military-civil fusion, and state direction has already reshaped procurement — DeepSeek V4 was reportedly withheld from US chipmakers and given to Huawei for exclusive early access, a signal of growing decoupling inside the Chinese model ecosystem itself.
"The U.S.-China AI model performance gap has effectively closed," Stanford's 2026 AI Index report concluded, "with a gap that fluctuated over the past year while remaining in the single digits."
UK Greens Want 38% of Big-Bank Profits
In London, the Green Party of England and Wales announced it would levy a 38% windfall tax on domestic profits above £800 million for the UK's biggest banks, aiming to raise at least £19 billion for small businesses. The money would be used to double the Employment Allowance for SMEs, taking up to £10,500 off their National Insurance bills — a design that targets the relief at roughly a million smaller companies.
Green Party leader Zack Polanski framed the proposal as a response to bank profits driven by monetary policy rather than improved services or products.
"While we are in a cost of living crisis, and with people unable to pay bills and small businesses being directly squeezed by interest rate rises which are expanding already obscene profits by the big banks, the Greens are calling for an end to rip-off Britain," Polanski said. "The large banks are cashing in on the backs of the small businesses who are the real innovators and creators, and profiteering from chaos and misery while ordinary people struggle."
The policy idea was floated by the campaign group Positive Money, which estimated it would raise £19 billion a year. The UK's four largest banks — HSBC, Barclays, Lloyds and NatWest — reported pre-tax profits of £45.9 billion for 2024, equal to around £650 per person in the UK, more than four times the profits recorded in 2020 before rates started rising, and nearly double the £25.6 billion average annual profits of 2018-21. The big four were on track to report another record, having posted profits of £24.1 billion in the first half of 2025 alone.
Polanski pointed to Barclays, where profits jumped in the six months to the end of June, lifting executive bonuses for the first half by nearly 30%, while the British Chambers of Commerce has described small and medium-sized businesses as facing a "cocktail of cost pressures." The Greens' core claim is that the biggest banks are making record profits not from improved services but from the high interest rates the Bank of England set to curb inflation.
The Political Economy of the Bank Levy
The banks are not starting from a zero-tax position. The bank Corporation Tax Surcharge currently stands at 3%, down from 8% after the previous government's review took effect in April 2023, with a £100 million allowance. The bank levy, introduced in 2011, taxes balance sheets rather than profits. When a similar windfall idea was raised last year, the trade body UK Finance said banks already pay both a corporation tax surcharge and a bank levy, and that a further tax would make Britain less internationally competitive and run counter to the government's aim of supporting the financial services sector.
The design precedent exists next door. The Energy Profits Levy introduced a 35% windfall tax on oil and gas companies, and the Greens' 38% figure is clearly calibrated against it. The Trades Union Congress has separately proposed raising the bank surcharge to 35%, estimating it would raise £12 billion in 2026-27 and £51.2 billion over four years, while a simpler reversal of the surcharge cut back to 8% would raise an additional £1.9 billion in 2026-27.
The Greens' argument rests on a specific claim: that record bank profits come from high interest rates set by the Bank of England to curb inflation, not from better products. That is a cyclical profit source. When rates fall, the surplus shrinks. A windfall tax designed to capture cyclical profits is, by construction, a bet on timing — and the timing risk cuts both ways. If the levy arrives after profits have already peaked, it raises less than advertised; if it arrives while profits are still elevated, it accelerates the political pressure on the governing party to do something visible before the next election.
The governing party has not embraced the proposal, and the Greens' limited parliamentary leverage makes the policy more a framing device than a near-term legislative probability. But framing devices have a way of becoming law when the political arithmetic changes.
What This Means: Two Surpluses, Two Regimes
The AI story is structural. The cost structure of frontier intelligence has changed permanently: efficiency gains, open-weight diffusion and domestic chip substitution mean the performance-per-dollar frontier is no longer a function of American hardware alone. That will not revert on its own. Even if US export controls tighten further, the routing-around behavior — third-country compute, domestic silicon, algorithmic efficiency — is now embedded in the Chinese AI ecosystem. The US is no longer defending an unbridgeable chasm; it is defending a lead measured in months and in cost differentials.
The bank-tax story is cyclical. Bank profits surged because the Bank of England held rates high; as inflation falls and rates normalize, those profits compress. The political pressure, however, may be more durable than the profits themselves. Once a windfall-tax precedent exists in a sector, it becomes a permanent feature of the political risk premium that investors price into UK financials, regardless of whether the specific levy passes.
The second-order implication is the asymmetry between the two. In AI, the US response — tighter controls, loophole closures — raises China's costs but does not restore the old lead, because the knowledge and the efficiency techniques have already diffused. In UK banking, the proposed response taxes a profit wave that is already receding, which means the revenue may arrive just as the profits that justify it disappear. One policy fights a structural shift with cyclical tools; the other fights a cyclical profit with a structural-looking tax.
There is a third-order angle the market has barely priced. If Chinese AI stays within months of the US frontier while costing a fraction as much, the global diffusion of capable models accelerates — and the pricing power of the US hyperscalers comes under pressure in every market outside the American regulatory perimeter. The winners of the AI race may not be the labs with the best benchmarks, but the ones that can monetize before the benchmarks become a commodity.
What to Watch
On AI, watch three signals. First, whether the Terminal-Bench and Intelligence Index rankings hold through the next release cycle — a Chinese model cracking the top three on Terminal-Bench would mark the end of the US performance lead, not just the cost lead. Second, Huawei's Ascend 950PR shipment volumes in the second half of 2026; if they materialize at scale, the hardware constraint loosens materially. Third, whether US authorities close the Southeast Asia data-center loophole — if they do, watch whether Chinese training runs slow or simply migrate again.
On the bank tax, watch the government's response and the autumn budget. The falsifying signal for the "cyclical profits" thesis is straightforward: if UK bank profits remain near record levels two years after the Bank of England cuts rates to neutral, the profits are structural, not windfall, and the tax debate changes shape entirely. The falsifying signal for the AI catch-up thesis is equally concrete: if the CFR's hardware-roadmap forecast holds and Huawei's next-generation chip is indeed less powerful than its current best, the performance convergence stalls and the US lead re-widens on the hardware axis even as the cost gap persists.
The base case: China's AI sector stays within months of the US frontier on benchmarks while undercutting on cost, and the UK bank-tax debate raises the political risk premium on British lenders without becoming law this cycle. The upside case for the Greens: a hung-parliament dynamic or a pre-election scramble forces a concession, and some version of the levy reaches the statute book. The downside case for China's catch-up: Washington closes the compute-access loophole and Huawei's roadmap disappoints, stretching the gap back out toward the CFR forecast.
The sharpest read of both stories is this: markets are being asked to price a world where technological leadership is measured in cost curves rather than crowns, and where the taxman arrives with a levy just as the windfall is about to blow itself out.
Explore more exclusive insights at nextfin.ai.
