But Its Externally Billed Line Covers a Fraction of What AWS Does — Disclosure first, and this entry needs it: I am an investor in ByteDance across rounds A through E (I wrote this in ENTRY 87, and it is in the site bio). That gives me a structural positive bias on this company, so in this entry I deliberately run the bear case longer than the bull and label the provenance of every number that flatters it. Read me at that discount. Now the number circulating in Chinese coverage, which is spliced: "ByteDance raised 2026 AI infrastructure capex 25% to ¥200bn, of which ¥85bn is for chips." Those two figures do not belong to the same budget. The ¥85bn comes from the Financial Times on 23 December 2025, attached to what was then a ¥160bn preliminary plan (about 53%). When the South China Morning Post reported the raise above ¥200bn on 9 May 2026, it published no chip line item at all — only that a proportionally larger budget had been allocated to domestic AI chips. The splice is traceable: Cailianshe via Tencent News/Chaoran Finance on 11 May 2026 printed both numbers in a single paragraph. That stitch also silently implies the chip share fell from 53% to 43%, which no outlet reported. The easier trap: ¥85bn appears in two unrelated contexts — the other is SCMP's 31 December 2025 figure for what ByteDance spent on Nvidia chips in 2025. And that same piece says 2026 Nvidia purchases would run about ¥100bn — ¥100bn for Nvidia alone exceeds the FT's ¥85bn for all AI processors, so the two reports are mutually irreconcilable and can only be presented as competing estimates, never as a breakdown of one budget; the ¥100bn also carries a condition, that the US permits H200 sales in China. Sources: FT (2025-12-23), SCMP (2025-12-31, 2026-05-09), Cailianshe via Tencent News/Chaoran Finance (2026-05-11, the splice), Bloomberg (2025-12-19, 2026-05-26, 2026-05-27), Sina Finance/Bianews (2025-12-27/28, Ascend orders), IDC China Enterprise MaaS Market Landscape, Omdia China AI Cloud Market 2025, Jiemian, LatePost (2025-06-05, 2026-06), 36Kr (2026-06-03), Huxiu (2026-08-06; its 2025 revenue figure relayed via GPLP 2026-07-28), Yicai/Securities Times (2026-04-20/21), Xinhua (2026-03-24), TMTPost/Xinlichang Pro (2026-05-22), Soochow Securities, and Volcano Engine's summer FORCE conference Q&A (2026-06-23). The ¥200bn number was also superseded eighteen days after it was reported, and more importantly every figure here is on a different basis. Fill in the baselines: 2025 AI infrastructure was ¥150bn on the FT's basis, and 2025 group-wide capex about $25bn on Bloomberg's. So the "surge" has to be computed separately — on the FT's own basis ¥150bn to ¥160bn is just +6.7%; SCMP's ¥200bn (about $30bn) against $25bn of group capex is about +20%; the genuinely large number is Bloomberg's 27 May report of internal discussions running as high as $70bn (roughly ¥475bn), with a further proposal for about $100bn in 2027 — though Bloomberg flagged these as preliminary and "could be very different" in execution. The committed figure remains ¥200bn+. Provenance matters too: the SCMP piece is a single-outlet exclusive sourced to two people familiar with the matter, Bloomberg's same-day item was a pickup rather than independent confirmation, and ByteDance discloses no capex and did not comment — treat it as an unconfirmed exclusive. A brokerage note claiming a further raise to ¥300bn is single-sourced; I don't cite it. And one thing Chinese reposts systematically dropped: SCMP gave two drivers, the second being rising memory-chip prices — which matters, because it means part of the increase buys price inflation rather than incremental compute. For scale: ByteDance's roughly $25bn of 2025 group capex compares with Tencent's ¥79.2bn and Alibaba's ¥126bn for the fiscal year ended March 2026 — it already outspends both listed rivals while disclosing nothing. What made me stop was not the size but where the money comes from — and I have to dismantle a framing I had in my own draft first. I was going to write about "three funding models": equity (ENTRY 120, two companies taking 43% of global venture funding in H1), debt (ENTRY 122, NVIDIA with six asset managers), and ByteDance's retained earnings. That taxonomy does not hold, and the counter-example is a sentence from my own ENTRY 117: "Amazon's TTM free cash flow is negative $7.6bn — $169bn of capex over the trailing twelve months ate the operating cash flow whole." Amazon's capex is also funded principally out of operating cash flow; the bonds only fill the negative. So the difference is degree, not kind, and the right frame is a coverage spectrum: capex divided by operating cash flow. ByteDance sits high on it — Bloomberg says it "will underwrite much of that spending through the roughly $50 billion of profit it earned in 2025." Amazon sits at about one or above (hence the negative FCF). Meta and the neoclouds sit below one and must top up with outside capital. I also retract a second line from my draft: I had written "no ByteDance bond issuance has been reported" as evidence — which is exactly the fallacy I dismantled yesterday in ENTRY 123 (absence of reports is not absence of events), and which my own ENTRY 122 framework contradicts, since off-balance-sheet structures, private credit and long-term lease or hosting commitments never show up in bond statistics in the first place. The accurate statement is narrower: the only visible financing trace today is profit, and that is "not observed," not "does not exist." What is genuinely hard on this spectrum is the threshold: if the $70bn figure lands, capex exceeds annual profit and ByteDance slides automatically into the tier that needs outside capital. Now the question I set myself: where does ByteDance sit on ENTRY 117's axis? Two honest caveats first. ENTRY 117 contained only a binary judgment — whether an externally billed revenue line exists — with no ratio threshold; the magnitude test is something I am adding today, not something the original said. And ENTRY 117 also stated that Meta's immediate trigger for being punished was the EPS miss, raised expense guidance and collapsing free cash flow; "externally billed" was the structural difference I drew, not the market's stated reason. On existence ByteDance passes, and cleanly: IDC's 49.5% MaaS share is defined in its own published methodology as "public-cloud model-invocation volume served by each cloud vendor to external customers, excluding own-business calls" — a genuinely external figure. On magnitude it fails, and the honest thing is to give three tiers rather than pick the flattering one: realized, MaaS revenue of about ¥1.5bn against 2025 AI infrastructure of ¥150bn is roughly 1%; on target, ¥15bn against ¥200bn is 7.5%; at the widest, all of Volcano Engine's revenue (a leaked range of ¥20-25bn) against ¥200bn is 10-12%. The comparable has to be the same coverage ratio, not a margin: AWS at a $169bn annualized run rate against Amazon's roughly $220bn of 2026 capex is about 77%. (AWS also runs a 39.4% operating margin — a different metric; don't mix them.) So ByteDance sits between Amazon and Meta, and closer to Meta than the token headlines suggest. But the two companies' "self-use" is not the same thing, and the difference cannot be written as "ByteDance has a money printer and Meta doesn't" — I wrote in ENTRY 117 myself that "Meta's advertising margins are not low." The verifiable difference is that ByteDance's AI self-consumption feeds directly into a distribution system already collecting ad revenue (Hongguo, Douyin, Ocean Engine), where every step of the loop has an existing billing surface, whereas Meta's AI self-consumption currently improves existing ranking and recommendation, with the incremental monetization still being proven. Three conflated metrics, and the first one is mine. First, "180 trillion tokens a day" is not an external-demand number: asked directly about the internal/external split, Tan Dai said "the 180 trillion is all Doubao model invocations, including internal and external, and including the Doubao app itself — the Doubao app is indeed a large share." I used this figure in ENTRY 105 (the 120-trillion version at the time); the framing I used then — in-house models as the group's ammunition depot — was right, but I did not label the basis, and I am labeling it now. Second, "Volcano Engine is China's number one AI cloud" depends on the basis: number one on IDC's invocation-volume measure, but number two on Omdia's AI cloud revenue measure (20.4% against Alibaba Cloud's 38.1%, with Alibaba alone exceeding numbers two through four combined); on the overall cloud market, Tan Dai said publicly in June 2025 that "Volcano Engine's share isn't even in the top three." Jiemian documented both companies running airport ads citing different research firms to each claim first place. Third, "China accounts for two-thirds of global model invocations" is wrong: Xinhua, citing OpenRouter for that week, gives a global total of 20.4 trillion tokens, with China's top-ten models at 7.359 trillion and the US top ten at 3.536 trillion — China is about 36% of the global total (roughly twice the US). Two-thirds is China's share of the China-plus-US subtotal. And the sample is only the top ten models on OpenRouter, a platform heavily skewed toward open weights. My own ENTRY 97 needs a correction, but not the one I thought while drafting. On 7 July I wrote: "more than half of Volcano Engine's MaaS revenue this year comes from Seedance alone (Tan Dai's own words)." Drafting this entry I was going to confess that I had understated it and it was really about 80% — that would have been a fake error: the original said "more than half," and 80% falls inside that. Nor is "80%" any outlet's figure; it is my own calculation, annualizing 36Kr's "over ¥1bn in a single month" and dividing by a leaked ¥15bn internal target — a numerator that is probably a peak month over a denominator that is an internal target, both soft. (LatePost separately puts Seedance's annualized revenue at about $2bn.) The real error is attribution: I put that figure in Tan Dai's name, and on 23 June he said publicly that every circulating Seedance revenue figure is wrong and inflated. I attached to him a number he has since disowned. That is the correction. Second, my "kept inside the house" framing is weakened: on the same day Tan Dai said "short drama is really just one deployment channel for Seedance, and in the long run it may be a small scenario," and that approaching half of Seedance usage is now overseas (Canva and multinationals). Note that is usage, not revenue, and those customers may carry lower ARPU. Even so, "kept inside to buy patience" is no longer its main form — I log this as partially weakened with the parent judgment still open, and the ledger entry for 97 was annotated before this entry went live. The bear case deserves real treatment, and the hardest item comes from ByteDance itself. One: share has not converted into revenue — Volcano Engine's 2025 revenue was about ¥20bn (Huxiu's figure, leaked; LatePost's separate version was a 2025 target above ¥25bn), "half of Tencent Cloud and a fifth of Alibaba Cloud," with MaaS only about 7% of that. Soochow Securities: it "needs to prove it can not only serve internal business but also establish independent platform value among external developers and enterprise customers." Two: Doubao monetization is close to zero — over 200 million daily users but under ¥1m in daily revenue, mostly e-commerce commissions, while as of May its daily compute cost ran into the tens of millions of yuan — "the compute cost of merely keeping Doubao running exceeds the entire operating cost of the Bilibili platform," against total user time under one-eighth of Bilibili's. (The response was launching Doubao Pro on 24 June at ¥68/200/500 tiers, ¥38 for students.) Three: unit economics are deteriorating, with an uglier signal attached — IDC's own 2026 forecast implies token consumption rising roughly 21x against about ¥186bn of revenue, meaning unit prices compress further, and Volcano Engine's survival condition has been summarized as "the rate at which compute costs fall must outrun the rate at which token prices fall"; the same analysis reads Volcano Engine bundling competitors' models (GLM-5.1, Kimi-K2.6) into its Agent Plan as an admission that its own Seed models no longer carry it alone — and that point attacks the quality of the externally billed line, which is worth more than the pricing half. Four: the content loop typecasts the ToB business as content-industry asset generation rather than a general enterprise AI substrate, with no hard non-entertainment exhibits at WAIC; on the demand side AI short dramas still have no breakout hit and "have never demonstrated the ability to replace live-action as a platform content pillar," alongside systematic likeness infringement. Five, the hardest, from ByteDance's own management: at the 6 August all-hands, "Volcano Engine MaaS, Doubao subscriptions and enterprise APIs all have to run positive cash flow eventually; burning money can't be the norm" — the implication being that today they don't; and the mandate given to the Seed foundation-model team at the same meeting was to "accept falling behind for a period and not look at short-term commercialization," an explicit acknowledgment that a large slice of the capex is not underwritten by near-term revenue. Three rulers for my own work. One: for capex-heavy companies, don't classify by equity/debt/profit — rank by coverage: capex divided by operating cash flow. That ratio determines whose instructions they follow in a drawdown; the portion below one is always paced by outside capital, and the larger that portion, the less the pace is theirs. Watch the threshold: once capex crosses annual operating cash flow, the company changes gear automatically. Two: test an "externally billed revenue line" on both existence and magnitude (the magnitude test is mine, added today, not ENTRY 117's original), and compare magnitude on the same coverage basis — 1% / 7.5% / 10-12% against 77% is not a small gap. Three: when you see "X trillion tokens a day," ask first whether it includes the company's own calls. The self-consumed portion generates no revenue for the group, only cost (internal transfer pricing is not external revenue); only the external-customer portion is a revenue line you can reconcile. Rollback: the ¥200bn/+25% figure comes from a single SCMP exclusive sourced to two people familiar, with Bloomberg's same-day item a pickup rather than independent confirmation and no ByteDance comment — treat as an unconfirmed exclusive; the ¥85bn belongs to the earlier ¥160bn plan (FT), is not part of the ¥200bn budget, and appears separately in an unrelated "2025 Nvidia purchases" context; the $70bn is an internal discussion range that Bloomberg explicitly flagged as preliminary; the brokerage claim of ¥300bn is single-sourced and not used here; SCMP's ¥100bn (Nvidia only, conditional on US approval of H200 sales in China) and the FT's ¥85bn (all AI processors) are mutually irreconcilable and presented as competing estimates; the "Ascend orders may exceed ¥40bn" figure is single-sourced in Chinese media (Sina Finance/Bianews); Volcano Engine revenue, MaaS revenue and Seedance revenue are all leaks or internal targets never disclosed by ByteDance, with Huxiu's ¥20bn relayed via GPLP on 2026-07-28; the internal/external revenue split has never been published and the circulating "about 70% internal" comes from an unnamed source, so I do not use it; the "net profit down more than 70%" figure is on an IFRS basis including fair-value changes in preferred shares and options, publicly rebutted by ByteDance VP Li Liang, so I use the operating-profit basis (his statement was that H2 2025 operating margin dipped slightly); the ">$50bn cash position" appears only in self-media estimates and is not cited; and I am an investor in ByteDance rounds A through E, so discount whatever in this entry favors it.
Falsifier: if within 12 months a confirmed debt or long-term lease/hosting commitment for AI infrastructure emerges (including off-balance-sheet structures, per ENTRY 122's framing rather than bond issuance alone), then "high coverage, low reliance on outside capital" is downgraded. Conversely, if within 12 months ByteDance or Volcano Engine discloses cloud or MaaS segment revenue publicly for the first time — whatever the number — then the "fails on magnitude" judgment can finally be tested properly; until then it is an inference built on leaks. Progress note on ENTRY 97: the "more than half from Seedance (Tan Dai's own words)" clause is corrected on attribution because he has publicly disowned all circulating figures; the "kept inside the house" clause is partially weakened by "approaching half of usage overseas" plus the company characterizing short drama as a small scenario; the parent judgment — SOTA-from-data multiplied by closed-loop distribution monetization — remains open.
This diary does not constitute promotion of, or an offer to subscribe for, any fund product.