Washington approved the chips. Beijing said no.

Last week this newsletter argued that the West is building smarter models while China builds habituated users, and that only one of those compounds. This week we price that claim. ByteDance's AI budget has been revised upward three times in six months, from $23 billion to above RMB 200 billion to a $70 billion ceiling now "under discussion." In the same period, reported net profit fell more than 70%. Both statements are true, and both are misleading. The number is a ratchet, not a plan. The profit collapse is mostly accounting. And the chip story keeps inverting: in August we wrote that the ban had already lost because everyone was using Chinese AI anyway. Now Washington has approved the chips, and Beijing is the one refusing them.

Answer first: ByteDance is deliberately subordinating near-term profit to infrastructure scale. The operational compression is real but modest. The money is not going to the free chatbot everyone watches; it is going to the enterprise inference toll booth and to silicon autonomy. That is where the value, and the risk, actually sit.

If you read only this far: anyone pricing Doubao's 382 million users is pricing distribution, not revenue. The asset to watch is Volcano Engine. The signal to read is China's refusal of the H200.

In December 2025, the Financial Times reported that ByteDance (字节跳动) planned to spend approximately $23 billion on artificial intelligence infrastructure in 2026. The figure was attributed to two people familiar with the matter. Anonymous sourcing is standard for this company, but what happened next is not: the number kept climbing.

By May 2026, the South China Morning Post, citing LatePost, reported the budget had risen above RMB 200 billion, or roughly $28.4 billion at prevailing exchange rates. Three weeks later, Bloomberg reported that a figure of up to $70 billion was "under discussion," subject to quarterly review. Bloomberg's report was corroborated by 36Kr.

The direction of revision is the only fact here. Any single number is provisional. What matters for allocators is that ByteDance has treated its AI budget as a ratchet, not a ceiling. Every revision has moved in one direction. This suggests the company is responding to competitive pressure, regulatory pressure, or both, rather than executing against a fixed plan.

The headline that accompanied these spending figures was stark: ByteDance's net profit had collapsed by more than 70% in 2025, from a trajectory that had the company on pace for roughly $50 billion for the full year after about $40 billion in the first three quarters, according to a December 2025 Bloomberg report. Full-year net profit came in closer to $9-10 billion, per Caixin and Yicai reporting in April 2026.

The apparent contradiction dissolves on closer inspection. Douyin Vice President Li Liang (李亮) confirmed on the record that the decline was driven primarily by IFRS fair-value accounting for preferred shares and employee stock options, not operating losses. The operational compression was real but modest: Li admitted that operating margins "dipped slightly" in the second half of the year.

The headline 70% decline is mostly accounting and timing. For allocators, the relevant question is which number they were underwriting. If you modeled $50 billion net income and got $10 billion, your model broke. If you modeled operating cash flow and found it compressed but intact, your model may still hold.

ByteDance's consumer-facing AI product, Doubao (豆包), reached 382.3 million monthly active users by June 2026, according to QuestMobile. That makes it the largest AI assistant in China by a wide margin. Its nearest domestic competitor, Alibaba's Tongyi Qianwen, reached 167 million MAU in the same period. DeepSeek stood at 129-130 million.

The scale is genuine. The monetization is not. According to a LatePost report cited by Guancha.cn, Doubao earned under RMB 1 million per day in the first half of 2026, primarily from e-commerce commissions on roughly RMB 10 million in daily gross merchandise value. Paid tiers, priced between RMB 68 and 500 per month, were not introduced until late June 2026. In May, when a subscription tease appeared, Doubao lost 6.1 million MAU.

36Kr reported explicitly that paid-user penetration is not a 2026 key performance indicator for Doubao. The company is treating the product as a distribution play, not a revenue center. This is the opposite of OpenAI's subscription-first model for ChatGPT. For allocators, Doubao's user count is a measure of distribution reach, not revenue potential. Anyone pricing the user base as if it were a monetizable asset is making a category error.

If Doubao is not the monetization story, what is? The answer lies in Volcano Engine (火山引擎), ByteDance's cloud and AI infrastructure unit. According to IDC data cited by Jiemian News, Volcano Engine holds approximately 49.5% of China's public-cloud model-as-a-service market by token volume. Alibaba Cloud is second at roughly 28%.

At the June 2026 FORCE conference, Volcano Engine President Tan Dai (谭待) disclosed that Doubao processes 180 trillion tokens per day, a tenfold increase over the prior year. The scale is enormous. The revenue, relative to capex, is not. Volcano Engine's 2026 MaaS revenue target is RMB 15 billion, or roughly $2.1 billion, a tenfold increase from approximately RMB 1.5 billion in 2025, per 36Kr. That target is still an order of magnitude below the capital expenditure figures being discussed.

The real monetization asset may be Seedance, ByteDance's video generation model. Multiple Chinese outlets report Seedance has reached approximately $2 billion in annual recurring revenue at gross margins near 70%, with monthly revenue exceeding RMB 1 billion, figures that Volcano Engine President Tan Dai publicly dismissed as "wrong and overestimated." Seedance revenue reportedly offsets a meaningful portion of Doubao's computing costs.

ByteDance is also shifting resources from consumer to enterprise. On July 30, 2026, the company announced that its Feishu productivity suite would be merged into Doubao, with the product team reporting to Zhao Qi and the go-to-market team reporting to Tan Dai at Volcano Engine. Liang Rubo (梁汝波), ByteDance's CEO, has publicly articulated a strategy of narrowing focus: the 2026 keyword inside the company is "Brave the Peak" (勇攀高峰), which means concentrating resources on AI model capability improvement and shrinking peripheral businesses. The public caution is equally instructive: Liang has also said "We are not yet ready," and noted that only a small group inside the company works on model training and that short-term returns will not be visible. The tension between the bullish internal keyword and the cautious public posture is itself an allocator signal.

The West's comparable spending is instructive. Meta has guided to $115-135 billion in 2026 capex, primarily for AI infrastructure. Amazon's guidance is approximately $200 billion. Alphabet's is roughly $185 billion. ByteDance's revised figures place it in the same conversation, though the company is smaller by revenue and lacks the cash-generation depth of its American peers. The $20 billion offshore syndicated loan reported by multiple outlets suggests the buildout is partly debt-funded.

The most consequential hardware story of 2026 has been widely misread in Western coverage. The standard narrative holds that Washington is denying China access to advanced AI chips. The actual story is more interesting.

In May 2026, the U.S. Commerce Department approved licenses for approximately ten Chinese companies, including ByteDance, to purchase up to 75,000 H200 units each. By June 29, zero chips had been delivered. Chinese buyers refused to accept the terms, which included a 25% revenue-sharing requirement.

This is a refusal, not a denial. Washington said yes. Beijing is saying no, and steering domestic demand toward Huawei Ascend (华为昇腾) chips, with the 950PR reportedly entering mass production in April 2026. For a year, the approved-but-undelivered H200 was 望梅止渴 (wàng méi zhǐ kě), gazing at plums to quench thirst: a promise of relief that never arrives. Beijing has stopped gazing and started digging the well, pairing the Ascend pivot with in-house silicon programs. ByteDance is reportedly trialing approximately 20,000 H200 units, worth roughly $400 million, while simultaneously building its own silicon. The company has assembled a 1,000-person self-developed chip team with four product lines, per 163.com and Securities Times reporting.

The geopolitical layer cuts both ways. Post-Manus, in April 2026, regulators reportedly began requiring ByteDance and Moonshot AI to obtain government approval before accepting U.S. capital, according to an analysis by international law firm Withersworldwide, which cites secondary news reporting. A separate report by 163.com corroborates the requirement for secondary-share transfers. The sourcing is qualified: these are reports of regulatory requirements, not published regulatory documents. The claim does not contradict the $20 billion offshore loan, which is debt, not equity.

For allocators, the signal is clear: China is willing to accept short-term performance penalties to build domestic chip ecosystem autonomy. Anyone modeling ByteDance's infrastructure costs on Nvidia list prices is using the wrong inputs.

Chinese technology history offers a playbook for cash-burn campaigns. Meituan's group-buying wars, Didi's ride-hailing subsidies, and the instant-retail race all ended in regulatory crackdowns and margin destruction. The State Administration for Market Regulation (SAMR) published ten typical enforcement cases (典型案例) of "involutionary competition" in February 2026, targeting network freight, mobile power, new energy vehicles, and solar photovoltaics.

AI infrastructure differs from these precedents in important ways. The assets are physical and depreciating, not digital and winner-take-all. The spending aligns with stated national priorities around technological self-sufficiency. The regulatory focus to date has been on consumer-facing price wars and platform abuse, not on B2B infrastructure investment.

Yet the precedent is a live tail risk, not a dismissal. The $20 billion loan signals that ByteDance is funding this buildout with borrowed capital, which means the cost of capital, the cost of regulatory delay, and the cost of competitive overhang all matter.

For allocators: Price the toll booth, not the traffic. Doubao's 382 million users are a distribution asset with no near-term monetization mandate. The revenue-generating infrastructure is Volcano Engine, and its RMB 15 billion target, even if achieved, is small relative to the capital being deployed. Seedance is the only asset currently generating material returns at high margins. Track Volcano's quarterly revenue against the capex ratchet; if it stalls while spending climbs, the funding gap and the debt load become the story.

For operators: Watch where the org chart moves, not where the press releases point. Feishu folded into Doubao, go-to-market moved under Volcano Engine, and the CEO says the company is "not yet ready." If you sell into China's AI economy, the buyer is the cloud arm, not the chatbot. If you compete with it, the consumer app is a subsidized moat, not the product.

For builders: A 1,000-person chip team, a deliberately small model-training group, and a consumer-to-enterprise resource shift tell you where the hard problems and the hiring budgets are. The frontier work at ByteDance this year is infrastructure and silicon, not another app.

The timeline signal: China is not waiting for Washington's permission to build its own stack. The domestic substitution push has moved from aspirational to operational. Anyone underwriting ByteDance's infrastructure costs should model Huawei Ascend, not Nvidia H200, as the marginal unit.

If this thesis is correct, four things reprice. Cost curves: if Ascend-scale inference lands at a structurally lower cost per token, AI pricing compresses globally, and margin assumptions built on Nvidia-era costs loosen everywhere. Valuation: Nvidia's China revenue assumptions and the listed proxies for Chinese inference share (Alibaba, Tencent), carry this story whether their models say so or not; ByteDance's own secondary-market marks above $550 billion embed a monetization path that has not started. Market structure: inference in China is becoming a state-aligned utility rather than a merchant cloud business, which changes what a 49.5% share is worth and to whom. Geopolitical exposure: export controls now bind in both directions, so "China risk" is no longer only what Washington does to Chinese firms; it is what Beijing declines to let them buy.

1. EAST: ByteDance discussing 2026 capex of up to $70 billion (Bloomberg, May 27, corroborated by 36Kr). Why it matters: the third upward revision in six months. The ratchet is the strategy.

2. EAST: H200 licenses approved, zero delivered, Beijing steering to Ascend (AllBright Law analysis). Why it matters: the export-control story inverted. Refusal, not denial.

3. EAST: Doubao launches paid tiers at RMB 68-500/month, with paid conversion explicitly not a 2026 KPI (36Kr). Why it matters: monetization deferred by design; 6.1 million users walked at the mere tease.

4. WEST: Hyperscaler capex guidance holds: Meta $115-135B, Amazon ~$200B, Alphabet ~$185B. Why it matters: ByteDance is converging on hyperscaler-scale commitment at a fraction of their cash generation.

5. EAST: Feishu merged into Doubao, GTM under Volcano Engine (Caixin, July 30). Why it matters: the consumer-to-enterprise pivot is now structural, not rhetorical.

Commandment Two: Own Your Space. ByteDance is applying it at industrial scale: 49.5% of China's inference market, the dominant consumer entry point, and a silicon roadmap are what owning your space looks like when the entry fee is $30 billion a year. The commandment cuts the other way for the reader. Most Western portfolios have no map of this space at all, and you cannot own a space you cannot see. Outside China, half the AI revolution is invisible.

The Ten Commandments of Investing, the Wall Street Journal and USA Today best-seller, is free for subscribers at workoptional.ai.

The deeper point: outside China, half the AI revolution is invisible. Western allocators tracking ChatGPT subscriber counts and Nvidia data center revenue are measuring one theater of a multi-theater war. ByteDance is building a parallel stack, on parallel silicon, under parallel regulatory logic. The numbers that reach English-language outlets are the ones that leaked. The ones that matter may not leak at all.

So: when the machine that spends your money is chosen, will you own the front door, or pay rent to whoever does?

Reply and tell me. I read every one.

San Eng

Disclaimer: East/West AIpha is for informational and educational purposes only and does not constitute financial, investment, legal, or tax advice. Nothing in this newsletter is a recommendation to buy or sell any security. The views expressed are our own, may be incomplete, and can change without notice. Do your own research and consult a licensed professional before making investment decisions. WorkOptional.ai, its writers, and affiliates may hold positions in companies or assets mentioned. Past performance is not indicative of future results.