Baidu's AI cloud tripled. The stock fell 13%.

AI & robotics investing — explained so you can actually act on it.

ISSUE 24 // Friday, 21 August 2026 // 7–8 min read

THE HOOK

On Tuesday morning Baidu reported GPU cloud revenue up 283% year over year, accelerating from 184% the quarter before, with AI crossing half of core revenue for the first time. On Tuesday evening Cerebras unveiled the fastest AI accelerator ever built and three brokers reiterated Buy.

Baidu closed down 13.1%. Cerebras closed down 4.3%.

Last week the hurdle rate was a document. This week we found out there's a second gate behind it, and the document doesn't open it.

WEEKLY SCOREBOARD

Prices are Thursday 20 August's close and session, except where marked — Baidu's move was Tuesday's earnings reaction and Cerebras' was Wednesday's launch response, both of which matter more than their Thursday marks. Macro: the Treasury said on Wednesday it would at least double repurchases of 10-, 20- and 30-year debt, and the thirty-year rallied ten basis points to 5.18%. By Thursday the ten-year had climbed back above where it sat before the announcement, and equities gave back the lot. Bessent told CNBC the operation could run larger than the $4bn announced. Brent traded near $91 as Trump threatened further action against Iran. The Philadelphia Fed's manufacturing index came in at 47.4 against a consensus of 25 — the highest since April 2021, which is not a reading that argues for lower rates.

Bottom line: Washington spent Wednesday trying to hold long-end yields down and couldn't manage 24 hours. The index gave back the rally, the AI complex went with it — and memory went up anyway.

TOP HEADLINES

1. Baidu’s AI business hit every number and the stock got destroyed.

AI Cloud Infra revenue RMB 7.3bn, up 50%. GPU cloud up 283%, a fourth straight triple-digit quarter and accelerating from 184%. Token usage on Qianfan up more than ninefold, embodied AI revenue up roughly sixfold, and AI-powered business crossing half of general business revenue for the first time. Then the other column: net income down 68% to RMB 2.3bn, gross margin at 39.0%, capex of RMB 11.4bn — nearly double the prior quarter — and free cash flow of minus RMB 7.95bn. The shares fell 13.1%. Morgan Stanley downgraded to Underweight the following morning and cut its target from $130 to $80. In June we set a test for this quarter — was Baidu’s AI Cloud growth accelerating or not? It is, emphatically, and it bought shareholders nothing.

2. Cerebras built the fastest chip on earth and got sold for it.

The CS-4 landed Tuesday at Supernova: three Wafer Scale Engines, twice the speed of the CS-3, ten times the throughput per watt, clearing 4,400 tokens per second per user on GPT-OSS-120B — up to thirty times a GPU’s per-user speed, with support for models beyond 50 trillion parameters. Needham and Rosenblatt held $300 targets. UBS called the launch the highlight of the event and the pullback an opportunity. Mizuho trimmed, explicitly on margins. The stock fell 4.3%. Read the guidance and the split makes sense: Q3 core revenue of $214–216m at a 38–40% gross margin, and an operating margin of minus 25% to minus 23%. The fastest inference in the world is currently sold at a loss.

3. Nvidia finally got into China and shipped 13% of what it’s allowed to.

ByteDance and Tencent have each taken delivery of roughly 10,000 H200s, per Wednesday’s FT. Washington licences each approved buyer for up to 75,000. Nvidia holds around 500,000 H200s built largely for Chinese customers, and Beijing is routing every order through the NDRC while pushing firms towards Huawei and Cambricon — some chips are being parked in Hong Kong, where power and data-centre capacity run out quickly. Two governments now restrict the same trade for opposite reasons and produce the same trickle. The licence stopped being the binding constraint some time ago. Nobody has said what replaced it.

4. The Treasury tried to hold yields down and lasted a day.

Wednesday morning: Washington announces it will at least double buybacks of 10-, 20- and 30-year paper. The thirty-year rallies ten basis points to 5.18%, equities lift, and for one session the AI complex has the falling discount rate it’s been asking for. It sold off anyway — Intel and AMD each down about 4%. Then Thursday reversed the lot, with the ten-year climbing back above its pre-announcement level and the S&P giving up 0.87%. Bessent went on television to say the operation could be bigger than the $4bn announced, which is not a sentence you deliver from a position of strength. The mechanism that’s supposed to rescue long-duration equities got tested twice this week and failed both times.

5. In Beijing: 3,000 robots, and one unfolded shirt.

The World Robot Conference opened Wednesday for five days with roughly 3,000 products and — despite the name — no visible foreign exhibitors. Unitree’s machines boxed, danced and played table tennis. UBTECH is selling an emotional-care humanoid to individual buyers at 168,000 yuan, about $24,000. At a neighbouring booth a robot spent several minutes trying to fold a shirt and failed. Physical AI is the one layer in our framework where capability genuinely is still the constraint, which makes it the one corner of this market where this week’s thesis doesn’t reach — and worth watching for exactly that reason.

OUR INVESTING ANGLE

Everyone’s asking what a company has to publish. The better question is what it has to publish it about.

The pattern: the disclosure standard held for a fortnight and then moved. Last week the market paid Sandisk 13.7% for a contract book and punished Applied Materials for a forecast. This week it read Baidu’s growth disclosure — 283%, accelerating, fourth straight triple-digit quarter — and took 13% off the stock anyway, because sitting beside it was a 39.0% gross margin and capex that nearly doubled in a quarter. The paperwork got filed. It wasn’t enough.

I want to be precise about what that is and isn’t. It’s an observed pattern across five sessions, not a demonstrated change in how this market prices things. Baidu had four other problems, and a market de-risking into a $5trn print would produce a tape that looks identical. The deep dive works through how much survives scrutiny; the short version is that the evidence is thinner than the story.

Who it would pay, if it holds. The Nasdaq fell 1% on Thursday and Micron rose 3.97%, holding the semiconductor complex positive on its own. Memory prices against a shortage that exists in fabs rather than in slide decks, and on a risk-off day that was worth seven points of relative performance — though the memory cycle explains that at least as well as my thesis does.

Who it hurts, specifically. Cerebras, at minus 24% operating margin on its own guidance. Baidu, already marked down and now carrying Wall Street’s first Underweight. AMD and Intel, both off around 4% on Wednesday on no company news at all, selling into a market everyone assumes is sold out. And the neoclouds — CoreWeave’s $104bn backlog was Issue 23’s exhibit for the disclosure thesis, and I’d note with some discomfort that its loss widened to $626m on $640m of interest in that same release. A backlog and a margin are different disclosures. We spent last week grading the first one.

What I can’t settle inside a Friday email is whether this is a regime or a fortnight — the same problem as last week, one gate further along. Both readings fit Thursday’s tape and imply opposite behaviour into Wednesday.

→ Read the deep dive: The second gate — what the market started charging for this week

THREE IDEAS TO RESEARCH THIS WEEKEND

Not recommendations — starting points for your own research. One continuation, one non-US, one I’m still arguing with myself about.

Idea 1 — Nvidia, and the one line that decides this.

Why now: 26 August is the cleanest test this thesis will get all year, and it lands five days from now. The case: the revenue beat is not in question — prediction markets put it at 94.5%, Jefferies and Stifel both model beat-and-raise, Morgan Stanley upgraded on Wednesday. A 94.5% probability isn’t a trade, it’s a settlement date. What’s genuinely unknown is the gross margin guide, and whether a beat delivered alongside a soft margin outlook still gets bought. The risk: one print, one company, sample size of one — and Nvidia is the least representative business in the sector precisely because it has never had to choose between growth and margin. Note also that CrowdStrike and MongoDB report the same evening, which will muddy the read considerably. Tripwire: Nvidia guides next-quarter gross margin below the level it guided in May and the stock falls on a revenue beat. That’s the second gate confirmed and Issue 23’s disclosure thesis superseded. Margin holds and the stock rallies, and both theses were a fortnight of tape. How to research: NVDA; SMH if you’d rather own the reaction than the event, bearing in mind it’s roughly a fifth Nvidia.

Idea 2 — Chinese domestic silicon, the accidental beneficiary.

Beijing is throttling H200 imports at roughly 13% of what Washington permits, and it isn’t doing that to protect anybody’s margin — it’s buying time for Huawei, Cambricon and the rest. Baidu’s own Kunlunxin chip broadened compatibility this quarter to Kimi K3, GLM 5.2, MiniMax M3 and Hunyuan 3, which is the unglamorous ecosystem work that decides whether domestic silicon becomes real rather than announced. Baidu also disclosed that domestic AI compute demand is growing against constrained supply — a shortage, in the one market where the shortage is policy-made. I keep coming back to that asymmetry. A deliberate constraint is more durable than a normal supply squeeze, not less, because it doesn’t resolve itself when capacity arrives. The risk: very little of this is cleanly investable from Europe, the listings are thin, and Beijing can reverse a policy-made shortage faster than it created one. Tripwire: the next H200 tranche. Deliveries landing near 10,000 per customer again while approvals stay open means the ceiling has stopped being regulatory and domestic capacity is winning on merit. Scale towards 75,000 and this idea is dead. How to research: BABA and 9984.T for the layer as we hold it; Cambricon and Hygon in Shanghai for the direct expression, with the access caveats that implies.

Idea 3 — Cerebras, where I genuinely don’t know.

Look, I’ve written this one three times. The bear read is in this issue’s angle: a 38–40% gross margin, minus 24% operating margin, and a market that has stopped paying for speed. The bull read is that this is one of the two or three best inference architectures in the world being marked down for spending money during a capacity build, which is precisely what you’d want it to be doing, with UBS and Needham both treating the pullback as the entry. Feldman is guiding to 600MW of capacity by end-2027 and a fourfold performance improvement. Customer concentration is the part that actually keeps me up — OpenAI, G42, MBZUAI and AWS is a very short list, and one renegotiation rewrites the model. It also IPO’d in May, so there is no multi-year record to check any of this against. Tripwire: Q3 results in November. Gross margin printing at or above the 38–40% guide with operating margin narrowing from minus 24% means the build-out is being paid for and my nervousness was misplaced. Below guide means the CS-4 is being sold into a price war it can’t win. How to research: CBRS directly; there’s no clean ETF expression, which is itself part of the risk.

AI INVESTMENT FRAMEWORK

Living portfolio framework by layer. Not financial advice — research starting points only.

Changes this week

● Infrastructure signal ↑ → ↔. Conviction HIGH holds. Last week the layer split between memory and equipment on disclosure. This week the seam moved: physical scarcity versus narrated scarcity. Thursday made it visible, with memory and merchant silicon moving in opposite directions on the same tape. The layer is no longer one trade, and NVDA carries the whole thing into Wednesday.

● Global signal ↑ → ↓. Conviction DEVELOPING holds, barely. Issue 23 named Baidu on Tuesday as this layer’s test. It delivered the growth and lost 13% anyway, then collected Wall Street’s first Underweight. That resolves the tripwire and confirms the growth is real. It also says growth is not what’s being paid for here. Five risk dots stay and I’d argue they’re light.

● Cybersecurity signal ↑ → ↔. Conviction MEDIUM holds. Issue 23 called this the one layer the thesis didn’t reach. That lasted six days — CrowdStrike sold off Thursday on its CTO’s departure, dragging the group. Not a margin event, but the immunity was overstated. CRWD reports 26 August alongside Nvidia.

● Platforms, Applications, Physical AI: no changes.

No ticker changes this week. Last week’s SAP-for-SKHY swap needs a quarter before it can be judged, and rotating the framework every Friday would make it a newsletter feature rather than a portfolio.

Per-layer notes

Changed layers are explained above — these notes cover what the table can’t.

Infrastructure Nvidia enters Wednesday carrying Issue 23’s three binaries plus a fourth this issue adds: the gross margin guide. Watch that line before the revenue line. Micron’s 208.9% year to date is now the layer’s return, not Nvidia’s.

Platforms Microsoft’s $678bn of contracted revenue in July remains the disclosure template and nothing this week tested it. The October follow-up quarter is still the gate on moving conviction.

Applications Unchanged and unhelpful. The above-consensus guide got you shot last week; this week nothing in the layer was asked to prove anything at all.

Physical AI Optimus serialised units at Fremont — Issue 20’s tripwire — has ten days left. Nothing in Beijing moves it either way. Separately, SpaceX’s second post-IPO unlock took effect Thursday with up to 319m restricted shares releasing, and the stock traded around $137.40; Issue 23 recorded the first overhang as cleared, and this is the second instalment of the same question.

Cybersecurity See above. Palo Alto reports 1 September, which is the layer’s real margin test rather than next Wednesday’s.

Global The tripwire resolved and the layer was marked down for passing it. Kunlunxin’s compatibility work is the thing to track here, not the headline revenue.

What we’re watching (next 6 weeks)

Beyond the window: SK Hynix Q3 in late October (Issue 23, Idea 2), and Cerebras Q3 alongside Applied’s fiscal Q1 in November (Idea 3, and Issue 23’s Idea 3).

Disclaimer: This newsletter is for informational and educational purposes only and does not constitute financial advice. iPrompt Signals is not a registered investment advisor. Always conduct your own research and consult a qualified financial professional before making investment decisions.

YOUR MOVE

Three rules the week leaves behind, not three facts you’ve now read four times:

1. Read the margin line before the growth line. If revenue is accelerating and gross margin is compressing, you haven’t found a growth company — you’ve found a company buying revenue. The market charged for the difference on Tuesday. Whether it keeps charging is Wednesday’s question.

2. Capex is a disclosure now. Go and find it. Baidu published RMB 11.4bn of it beside a 283% growth number and the market read the smaller figure. Spending has stopped being the invisible half of the story.

3. A backlog proves demand exists. It doesn’t prove demand is profitable. Last week’s lesson still holds — it just stopped being sufficient on its own. Both numbers, or neither.

This week’s one ask. Take one AI position you own and pull its gross margin for the last four quarters, alongside revenue growth for the same four. Put them side by side. If margin is falling while revenue rises, you own a Baidu — which may well be fine, but you should know it deliberately rather than discover it on an earnings morning. Reply with the ticker and both trend lines; I’ll run the three most interesting next Friday.

Optional depth: this week’s deep dive turns that into a repeatable four-quarter screen, and works through the two cases where a falling margin is the right answer.

Stay curious — and stay qualified.

— R. Lauritsen

Editor, iPrompt Signals

Know someone building an AI position? Forward this.

P.S. — Issue 22’s Baidu tripwire resolves cleanly. AI Cloud growth was accelerating, GPU cloud nearly quadrupled, and the answer was an unambiguous yes. The stock fell 13% the same morning. I set that tripwire in June and it did exactly what a tripwire is meant to do: gave me a dated, checkable question and a real answer on the day. The answer turned out to be worth nothing, because I’d asked about the wrong line in the accounts. Tripwires keep you honest about what you predicted. They can’t keep you honest about what you chose to measure.

QUICK GLOSSARY

Capex (capital expenditure) — Money spent on buildings, chips and equipment before any revenue exists. Doesn’t show up in profit; shows up in cash. Baidu’s was RMB 11.4bn in Q2.

Free cash flow — Cash from operations minus capex. Baidu’s ran to minus RMB 7.95bn.

Gross margin — What’s left of revenue after the direct cost of producing the thing. Cerebras guides 38–40%; Baidu printed 39.0%.

Operating margin — What’s left after staff, R&D and overheads too. Cerebras guides minus 25% to minus 23%.

NDRC — China’s National Development and Reform Commission, which must approve every H200-based order individually.

Rack-scale — A whole server rack sold as a single product rather than chip by chip. The CS-4 is Cerebras’ first.

Tokens per second per user — How fast a model answers one person, rather than how many people it serves at once. The CS-4 claims 4,400 on GPT-OSS-120B.

Treasury buyback — The government repurchasing its own outstanding debt to support prices and hold yields down. Announced Wednesday, unwound by Thursday.