In partnership with
Welcome to the iPrompt Newsletter
Friday Edition
Sponsor:
4x your communication output. Same quality. No burnout.
The bottleneck isn't what you want to say — it's how long it takes to type it. Wispr Flow removes the bottleneck.
Speak naturally and get polished, send-ready text for executive summaries, client updates, board recaps, investor notes, or just the 30 Slack messages you're behind on. Flow strips filler, formats numbers and lists, and preserves your tone.
Used by teams at OpenAI, Vercel, and Clay. 89% of messages sent with zero edits. Works in every app on Mac, Windows, and iPhone.
iPrompt Signals
AI & robotics investing — explained so you can actually act on it.
ISSUE 23 // Friday, 14 August 2026 // 7–8 min read THE HOOK
On Thursday morning Sandisk stood up in New York and showed investors $93.9 billion of signed contracts. On Thursday evening Applied Materials reported a record quarter and guided October revenue seven per cent above the Street.
Sandisk closed up 13.7%. Applied Materials fell, then fell again on Friday.
Last week the hurdle rate moved. This week we found out it isn't a price — it's a document.
WHAT TO DO WITH IT Bull case: the side that can publish a contract book — memory, the neoclouds, anyone with a remaining performance obligation line in the accounts. Bear case: Nvidia reports on 26 August without one and could re-rate the whole cohort anyway. Named risk: wafer-fab equipment, which has no document to publish and just found out what that costs. Tickers, both sides, below. |
WEEKLY SCOREBOARD
TICKER | THU CLOSE | THU % | WHAT HAPPENED |
SNDK | $1,528.11 | +13.7% | Investor Day: $93.9bn of contracted revenue from eight customers, backed by $16.5bn of guarantees. Eight days after falling ~8% on a record quarter. |
AMAT | $534.54 | −2.5% | Record Q3, October guide ~7% above consensus, "unprecedented visibility." Sold anyway — and roughly 5% lower again into Friday's open. |
NVDA | $225.30 | +0.5% | +2.9% on last Thursday's $218.99, despite Monday's 2.4% drop on an FT report of a $500bn AI financing package. Market cap $5.45trn. 26 August next. |
MU | $949.83 | +4.2% | Dragged up in Sandisk's wake. Already sits on 16 Strategic Customer Agreements and ~$100bn of contracted revenue — disclosed in June, unrewarded since. |
S&P 500 | ≈7,800 | +0.6% | Record close, through 7,800 for the first time (intraday 7,816.70) on a soft PPI print. |
NASDAQ | 26,803.03 | +0.8% | Third straight weekly gain, led by memory and the AI-infrastructure complex. |
VIX | 14.63 | +0.6% | Touched a new 2026 low intraday before bouncing. Nobody is paying for protection into Nvidia. |
Prices are Thursday 13 August's close and session. VIX = the "fear gauge" — expected S&P 500 swings over the next 30 days. Above 25 = real anxiety; 14 = a market that has stopped imagining downside. Macro: in-line CPI and a soft PPI cut September rate-hike odds to roughly 30% from 45%, which is what put the index through 7,800. Friday's retail sales then fell 0.6%, the worst in over a year. The rate relief arrived as weakness — that bill comes later.
DRIFT STAT WATCH The −13.4% cohort drift figure stays our standing yardstick to Micron's late-September print. No update this week — CNBC hasn't refreshed the dataset. On the individual tape: Sandisk left the punished column, Applied Materials joined it. |
Bottom line: Two record quarters in the same memory complex eight days apart. One got sold, one got paid — and the difference wasn't the numbers, it was the paperwork.
TOP HEADLINES
1. Sandisk answered the hurdle rate with a document, not a quarter.
On 5 August Sandisk reported record revenue, an 84.6% gross margin and a large beat — and fell about 8%. On Thursday it held its Investor Day and disclosed $93.9bn of minimum contracted revenue across ten agreements with eight customers, three of them US hyperscalers, weighted-average duration over four years, backed by $16.5bn of financial guarantees. The book covers more than half of fiscal 2027 bit shipments. The stock closed up 13.67% at $1,528.11.
2. Applied Materials ran the control experiment on the same day.
This is what the other side looks like. Record fiscal Q3: revenue $9.12bn, up 25%; adjusted EPS $3.50 against $3.39 expected. Then a guide most companies would kill for — October revenue of $10.25bn against a $9.55bn consensus — that is, seven per cent above what analysts had modelled — with EPS guided well clear too. Management called its demand visibility unprecedented, citing customer conversations running out to 2030. The stock fell 2.5% on Thursday and about 5% more into Friday.
🌱 NEW TO INVESTING? HERE'S WHAT THIS MEANS Guidance and a backlog are not the same thing. Guidance is management's forecast — a promise about next quarter that costs nothing to make. A backlog, or "remaining performance obligation", is money customers have already committed under signed contracts, and it sits in the accounts. Applied Materials gave the market a very good forecast. Sandisk gave it $93.9bn of signed orders. |
3. CoreWeave's number was the backlog, not the revenue.
Revenue $2.58bn, up 112%. Loss widened to $626m on $640m of interest. Ordinarily that gets executed. Instead the stock rose 21% on Wednesday, because the backlog hit $104.2bn — up 246%, with contracts beyond 48 months rising from 10% to 21% of the mix and over $25bn signed since quarter-end. The market read the release and graded the contract book.
4. DeepSeek quadrupled its prices. Read that as a demand print.
The cheapest frontier lab on earth is getting expensive. From Sunday, V4-Flash output goes from $0.28 to $1.32 per million tokens at peak, V4-Pro from $0.87 to $3.96 — the stated aim being to allocate resources more reasonably, which is a polite way of saying capacity is congested. Note the form it takes: a published price is a disclosure too, and it's the first hard number anyone has put on inference scarcity.
🌱 NEW TO INVESTING? HERE'S WHAT THIS MEANS AI companies charge by the "token" — roughly a fragment of a word. For two years those prices only fell, which is why so many people assumed AI would get cheaper forever. A fourfold rise, with peak and off-peak rates like an electricity tariff, says something different: the compute behind the model is scarce enough to ration. That scarcity is what memory and infrastructure shareholders own. |
5. OpenAI let its revenue slip. That's not a leak — it's a rehearsal.
Bloomberg reported on Thursday that OpenAI is running at more than $40bn of annualised revenue, roughly double its end-2025 rate, ahead of a confidential SEC filing lodged in June. Rank that against everything above and it sits at the bottom: a run rate is unaudited, uncontracted and attributed to people familiar. It's the weakest form of disclosure there is — which is precisely why it leaked rather than being published.
OUR INVESTING ANGLE
Everyone's asking whether the AI selloff is over. The better question is what a company has to publish to end its own.
The thesis: the hurdle rate isn't a valuation ceiling — it's a disclosure standard. That distinction matters enormously, because a multiple is something the market does to you, while a disclosure is something management chooses. Sandisk proved it inside eight trading days: same company, same NAND market, same demand curve. What changed was that on 5 August it offered a record quarter, and on 13 August it offered a contract book. The market paid 13.7% for the second one.
Who this pays. The companies that can convert assumed demand into contracted demand and then show the paper: Sandisk now, CoreWeave already, Microsoft since July — and Micron, which published sixteen Strategic Customer Agreements and roughly $100bn of minimum contracted revenue back in June and has not been re-rated for it.
Who it hurts, specifically. Wafer-fab equipment. Applied Materials, Lam Research and KLA sell tools against rolling forecasts and purchase orders, so there is no contract book to publish and "unprecedented visibility" is the strongest claim available to them. This week we learnt exactly what that claim is worth against a document. Add Cisco, which guided revenue above consensus on Wednesday and fell 8.8%, and AMD, which sells into a market everybody assumes is sold out. An assumption is not a contract.
What I can't settle inside a Friday email is the bit that decides whether any of this repeats: is the market paying for the disclosure itself, or simply repricing economics it couldn't see before? Those look identical on Thursday's tape and imply opposite behaviour from here. That question, the five-part test I now use to separate a contract book that counts from one that's marketing, and the two ways this ends are the whole of this week's deep dive.
→ Read the deep dive: The disclosure premium — what a contract book is worth in 2026
⚠️ WHAT COULD GO WRONG? (THE BEAR CASE) 1. A floor price is also a ceiling. [attacks the thesis] Sandisk traded upside for certainty — fixed pricing near term, collars later. JPMorgan reckons gross margin holds near 80% even at floor prices, but if NAND keeps running, that book caps the very upside that made the stock a 450% year. Getting paid for a floor only looks clever if the floor binds. 2. Nvidia doesn't need the trick. [attacks this week's expression] On 26 August the largest company in the world reports without a published contract book and could re-rate the entire cohort in a single session. If a beat finally gets bought there, the market has shown that at the top of the stack results still work unaided — and this was a fortnight of tape, not a standard. The most likely way to be wrong here is to be right about the mid-caps and early on everything else. 3. A backlog is only as good as who signed it. [attacks the thesis] CoreWeave's $104bn rests on counterparties that are themselves loss-making and levered — and CoreWeave's own loss widened to $626m on $640m of interest. Sandisk's guarantees cover $16.5bn of $93.9bn, about 17%. The rest is a promise in a nicer font. Size your position for the possibility that this is a two-week fashion in disclosure rather than a change in what the market values. |
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 — Micron, the disclosure that's already been made.
Why now: Sandisk was paid 13.7% for publishing a contract book. Micron published one in June and got nothing for it. The case: sixteen Strategic Customer Agreements, roughly $100bn of minimum contracted revenue across fourteen of them, running to 2030 and covering about 20% of DRAM volume and a third of NAND. If the market has started grading paperwork, Micron's is already filed. The risk: DRAM isn't NAND — Micron's leverage runs through HBM, exactly where Nvidia's reported reduced-memory Rubin Ultra designs would bite. Tripwire: at Micron's fiscal Q4 print in late September, does a quantified contracted-revenue figure appear on a slide the way Sandisk's did? Quantified and the comparison is direct. Described but not numbered and the market has nothing to grade. How to research: MU; SMH or the Roundhill Memory ETF (DRAM) for the complex.
Idea 2 — SK Hynix, the largest unpublished contract book in memory.
So who's third? Sandisk published, Micron published, and Thursday dragged SKHY up alongside MU before Friday added a semiconductor-led rally in Seoul. SK Hynix is the HBM leader, listed ADRs on the Nasdaq in July raising about $29bn, and its 2026 supply is already sold out under multi-year agreements — it has simply never put a contract-value number in front of investors in the Sandisk format. I keep coming back to that gap. The disclosure exists commercially. It just hasn't been made. The risk: this is also the company whose comments about slowing its AI memory business helped trigger one of the Kospi's worst sessions on record, and the memory oligopoly is facing a price-fixing suit. Tripwire: SK Hynix's Q3 results in late October. A disclosed contracted-revenue or minimum-volume figure makes three of three memory makers publishing — an industry standard rather than a Sandisk quirk, and the strongest confirmation this thesis could get. How to research: SKHY for the US line, 000660.KS in Seoul.
Idea 3 — The equipment makers, on the wrong side of it.
Look, I'm uneasy pointing at a group whose fundamentals just improved. Lam raised its 2026 wafer-fab equipment outlook to $150bn from $140bn this week, and neither LRCX nor KLAC moved much on Applied's fall — the market read that selloff as company-specific. But if grading really has shifted to disclosure, equipment is structurally disadvantaged: you sell tools against rolling forecasts, so there is no contract book to publish, and every strong quarter arrives wrapped in the same soft visibility language that just got Applied sold twice in two sessions. Permanent discount or mispricing? I don't know yet. Tripwire: Applied's fiscal Q1 print in November. A quantified committed-order figure it has never previously disclosed = the arms race is real and the discount closes. Another record quarter, another "unprecedented visibility", another selloff = it's structural. How to research: AMAT, LRCX, KLAC; SMH if you'd rather own the complex than the question.
AI INVESTMENT FRAMEWORK
Living portfolio framework by layer. Not financial advice — research starting points only.
Changes this week
● Infrastructure signal ↑↑ → ↑. Not a conviction cut — HIGH holds. But the layer has split internally: memory can publish a contract book, equipment cannot, and this week the market paid sixteen points of spread for the difference. Entry-price discipline becomes entry-disclosure discipline.
● Applications signal ↔ → ↓. Cisco guided revenue above consensus and fell 8.8%. Globant fell 12.6% and cut its full-year outlook. The one Applications name that rose meaningfully was Workday, up 17.8% on takeover speculation. A layer being valued by acquirers rather than by operators is not a healthy layer.
● Global: SAP out, SKHY in. SAP has been a diversifier without an AI-specific catalyst for three quarters. SK Hynix gives the layer direct HBM exposure, a US-listed line since July and a dated disclosure catalyst in October. Re-entry trigger for SAP: a quantified AI-related backlog or cloud-ACV disclosure at its next capital markets day. Signal ↔ → ↑; conviction stays DEVELOPING — Baidu on Tuesday is the test.
● Platforms, Physical AI, Cybersecurity: no changes.
LAYER | TICKERS | CONVICTION | RISK | SIZING |
INFRASTRUCTURE | NVDA, MU, INTC | HIGH ↑ | ●●●○○ | 15–20% |
PLATFORMS | MSFT, AMZN, GOOGL | MEDIUM ↑↑ | ●●●○○ | 5–10% |
APPLICATIONS | PLTR, CRM, NOW | MEDIUM ↓ | ●●●●○ | 5–10% |
PHYSICAL AI | BOTZ, ISRG, TSLA | DEVELOPING ↑ | ●●●●○ | 5–10% |
CYBERSECURITY | CRWD, PANW, ZS | MEDIUM ↑ | ●●●○○ | 5–10% |
GLOBAL | BABA, 9984.T, SKHY | DEVELOPING ↑ | ●●●●● | 5% |
Per-layer notes
Changed layers are explained above — these notes cover what the table can't.
Infrastructure NVDA enters a 26 August print carrying three binaries. Worth noting on INTC: its market value has grown 474% over twelve months against Nvidia's 24%. This layer's returns are no longer one name's story.
Platforms The layer that set the standard. Microsoft's $678bn of contracted revenue in July is the template every company in this issue is now measured against. Conviction stays MEDIUM until the October follow-up quarter lands.
Applications The guide still gets you shot. Now the above-consensus guide gets you shot too.
Physical AI SpaceX cleared its lockup and ran roughly 35% in eight sessions back through its $135 June listing price — Issue 22's tripwire resolves as "overhang cleared", with the caveat in this issue's P.S. Optimus serialised units at Fremont, Issue 20's tripwire, still unresolved with seventeen days on the clock.
Cybersecurity Still bought from operating budgets, still untouched by the memory tax — and still not required to publish anything to prove it. The one layer this week's thesis doesn't reach.
Global Five risk dots stay. Baidu reports Tuesday pre-market; DeepSeek's fourfold price rise lands the day before.
What we're watching (next 6 weeks)
DATE | EVENT | QUESTION TO TRACK |
16 Aug | DeepSeek V4 peak pricing live | Do Western API providers follow the rise — or take the share? |
18 Aug | Baidu Q2 earnings, pre-market | AI Cloud growth: accelerating or not? Issue 22's tripwire, live. |
26 Aug | Nvidia Q2 earnings | Three binaries: the $250bn OpenAI guarantee in the filing, reduced-memory Rubin Ultra confirmed or denied, and whether a beat finally gets bought. |
31 Aug | Optimus serialised units, Fremont | Issue 20's tripwire, still unresolved. None by month-end and Japanese robotics graduates to primary. |
Late Sept | Micron fiscal Q4 earnings | Does a quantified contracted-revenue number appear on a slide? Idea 1's tripwire — and the drift stat resolves here. |
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. Check what a company can publish before you check what it earned. If there's no contracted-revenue line available to management, a record quarter is the ceiling of its good news, not the floor.
2. An above-consensus guide is no longer a catalyst. It's the entry fee. Price the guide before the print, and assume the beat is already owned.
3. A widening loss is survivable; an unquantified order book isn't. The market spent this week forgiving one and punishing the other.
This week's one ask. Take a single AI position you own and go find its remaining performance obligation, backlog or contracted revenue in the last quarterly filing or investor deck. If the company doesn't publish one, write down what it publishes instead — then ask yourself what the market is actually being given to grade. Reply with the ticker and the number; I'll run the three most interesting next Friday.
Optional depth: this week's deep dive turns that exercise into a five-column test you can reuse on any company, in any sector.
🌱 SHORT TAKE (FOR THE BROAD-EXPOSURE READER) If working out which companies publish a contract book sounds like homework you won't do, own the complex instead of the question: SMH or SOXX for semiconductors broadly, or the Roundhill Memory ETF (DRAM) for the part of this week's story that actually got paid. The volatility is real — Sandisk alone swung 13% in a single session. Not a recommendation, a starting point. |
Stay curious — and stay qualified.
— R. Lauritsen
Editor, iPrompt Signals
Know someone building an AI position? Forward this — they'll thank you by Friday.
P.S. — Issue 22's SpaceX tripwire resolves: 900 million shares came unlocked and the stock ran about 35% in eight sessions back through its listing price. The overhang cleared. What I didn't say last week is that short sellers being squeezed 40% off the low looks identical to demand on a chart. Tripwires tell you what happened. They don't tell you why.
QUICK GLOSSARY
Backlog — Orders signed but not yet delivered. CoreWeave's is $104bn against roughly $13bn of expected 2026 revenue.
Floor price — The minimum a customer pays under a long-term contract, whatever the spot market does. Protection on the way down; a cap on the way up.
New Business Model (NBM) agreement — Sandisk's name for its multi-year, fixed-price supply contracts. Ten of them, eight customers, $93.9bn.
Peak/off-peak pricing — Charging more when demand is highest, like an electricity tariff. DeepSeek adopts it on 16 August.
Remaining performance obligation (RPO) — Contracted revenue not yet recognised. The audited version of a backlog.
Strategic Customer Agreement (SCA) — Micron's equivalent of Sandisk's NBM. Sixteen signed, roughly $100bn.
Wafer fab equipment (WFE) — The machines that make chips. Applied, Lam and KLA sell it — against forecasts, not contracts.
iPrompt Signals Published Fridays by FrontWave Media Ltd · Limassol, Cyprus |
Sponsor:
22 ChatGPT Agents Built for Every Marketing Job
Most marketers use ChatGPT to do general research and then call it an AI strategy. The ones outperforming them are deploying specialized agents built for specific jobs.
We put together 22 plug-and-play ChatGPT marketing agents that handle the work eating your week, each with built-in instructions and structured outputs ready to go in under 5 minutes.
Subscribe to Marketing Against the Grain and get all 22 free.
Inside you'll find:
Competitive intelligence agent that visits competitor websites and builds detailed comparison matrices automatically
Customer feedback analyzer that ranks improvement opportunities by business impact
Social listening specialist that monitors brand mentions and flags reputation risks before they escalate
Campaign optimization agents that handle attribution analysis and surface what is actually driving results
Your competitors are already running agents like these.
Get 22 ChatGPT Marketing Agents free when you subscribe to Marketing Against the Grain today.