The Constraint Chain, the Financing Loop, and the Efficiency Shock

Day three of Goldman's technology conference was the day the biggest company in the industry stood up and put a number on the next twelve months. Nvidia said it can grow roughly 70 per cent again, and that if it could build everything customers wanted, growth would be over 100 per cent. Oracle then reported after the close with a backlog of 664 billion dollars. Taiwan Semiconductor said its August sales were 53 per cent higher than a year ago.

And the market fell for the fourth day in a row. The reason is not in any of those numbers. It is in the cost of money: producer-price inflation came in at 5.4 per cent, the 30-year Treasury yield hit its highest level since 2007, oil went through 100 dollars because of the fighting around the Strait of Hormuz, and the odds of the Federal Reserve raising rates next week moved to about 70 per cent.

One more thing happened that nobody had on the card. A Chinese lab, DeepSeek, released a model that needs about a quarter of the memory per unit of work that its previous model needed. Memory shares fell. That matters because 24 hours earlier, on day two, the memory industry had spent the morning telling investors that prices are now protected by long-term contracts through 2030.

Three days, one message, one contradiction

The tape and the macro regime, 10 September

Start with the price action, because it is the part of day three that contradicts everything the executives said. Four consecutive down days into a hot producer-price print, a two-decade high in long-bond yields, and oil through 100 dollars.

1.1 What the curve did

Fifty-two-week highs across the curve; the long bond is at its highest since 2007. Bear flattening at the front, driven by the hike, on top of a long end already repricing term premium. Fed funds currently 3.50 to 3.75 per cent. The FOMC meets 15 to 16 September.

August PPI printed plus 0.4 per cent month on month, in line, and plus 5.4 per cent year on year against 5.3 per cent consensus. Core PPI 4.6 per cent. The miss was in the annual rate, which is the number that moves hike odds. CME FedWatch odds of a 25 basis point increase moved to roughly 70 per cent from roughly 56 per cent before the print. Chair Warsh's 28 August remarks, “work to do” on price stability, remain the anchor; Hammack pushing to act, Waller and Williams data-dependent.

1.2 What else moved

Oil. WTI closed around 100 to 102.5 dollars, up 4.2 per cent or 6.43 dollars depending on source, and Brent around 105 to 108, on escalating US-Iran attacks on shipping. OPEC lowered its 2026 demand-growth forecast into the move, which makes the rally supply-driven, not demand-driven. That is the worse kind for equity multiples.

Volatility. VIX broke above 17 after 28 consecutive sessions inside 14 to 17. The volatility regime changed on day three, quietly, and almost nobody wrote about it. Gold fell 53.40 dollars to 4,407.30. Freeport fell 7.3 per cent, Intel 5.7 per cent after a multi-session rally, and the Russell 2000 1.0 per cent, the fourth day of small-cap underperformance.

TSMC. August revenue was NT$514.806bn, up 53.3 per cent year on year and 10.1 per cent month on month on the July record. January to August cumulative NT$3.38687tn, up 39.3 per cent. The 2026 capex guidance stands at a record 60 to 64 billion dollars.

The read-through is a duration read-through

AI infrastructure equity is the longest-duration cash-flow profile in the public market: the capex is now, the free cash flow is 2028 and beyond, and the terminal value is most of the price. Marking that against a 5.37 per cent 30-year is mechanically hostile regardless of how good the operating news is.

Day three is the cleanest demonstration of the bear-steepener thesis available: fundamentals up, multiples down, and the highest-duration names taking the most damage. Nothing in the tape invalidates a single company thesis. Everything in the tape argues that the sequencing of the next twelve months is set by the discount rate, not by the order book. Energy exposure is the only part of the complex that gets paid by the Hormuz premium. Everything else pays for it.

Nvidia: the growth number, and what sits under it

We are tracking every single gigawatt of land, power and shell around the world, literally everything on the planet.

Jensen Huang, Chief Executive Officer, Nvidia

We put in one and 100 comes back in.

Huang, on the circular-financing question

Two sentences that describe the same business model from opposite ends: Nvidia as the industry's supply-chain planner, and Nvidia as the industry's balance sheet. The session was the reason the day was on the calendar, and it delivered more than a reiteration. Work through it in order, because the conclusion only holds if each step does.

2.1 The number

I think we could grow 70% year over year. We're confident about that.

Jensen Huang, Nvidia

Same figure as the 26 August FQ2 call, now said unprompted at a conference three weeks later, which upgrades it from a guide to a commitment. Implied revenue of roughly 680 billion dollars next fiscal year against roughly 400 billion in the current one, which is journalist arithmetic rather than a company figure and should be treated as an estimate.

Unconstrained demand growth is above 100 per cent. Huang said demand is growing faster than supply and that Nvidia has roughly one year to expand supply to meet it. The 70 per cent is therefore a supply number, not a demand number. That distinction is the most important thing said at the conference and it was widely reported as though it were a demand forecast. Global AI infrastructure spend of 3 to 4 trillion dollars by 2030, on the end of Moore's law plus hardware-software co-design.

2.2 The constraint list

This is the most useful thing Huang said, and it is not the growth number. Asked where the bottleneck is, he did not answer with one item. He answered with a chain.

Day two said the constraint moved from silicon to power. Day three says it did not move; it multiplied. A single bottleneck is a trade: find it, own it, wait for the shortage rent. A serial chain of eight constraints is a different structure. Each link can bind in turn, so the shortage rent rotates, and the clearing rate of the whole queue is set by whichever link is worst at any moment.

One: breadth across the chain dominates concentration in the fashionable bottleneck. Owning the power name in a quarter when connectors bind earns nothing.

Two: the marginal dollar belongs with links that have both price-setting power and the balance sheet to add capacity without the capital markets. Pricing power without funding capacity is a trap in a 5.37 per cent long-end regime, which is the whole of Section 3.

2.3 The circular-financing answer

Huang's frame: Nvidia only invests behind contracted revenue, citing roughly 100 billion dollars of contracted revenue ahead of the investments. Neoclouds were named explicitly as the regional capacity layer beyond the hyperscalers: CoreWeave, Nebius, Nscale, Lambda, Firmus. Nebius was separately initiated at Buy by Truist the same day.

Nvidia runs every model. Every single lab can use us.

Huang, framing Nvidia as a foundational platform of the AI ecosystem

2.4 The two new growth legs

Cybersecurity, named as a likely next growth engine for AI, with CrowdStrike, Cisco and Palantir cited as partners. Physical AI, meaning robotics and autonomous driving, with meaningful advances expected over two to three years. Drug discovery, warehouse robots and AI-native telecom networks also named.

The translation: Nvidia is telling investors where the next wave of chip demand comes from once the chatbot and coding buildout is spoken for. It named security software and robots. For anyone already treating cyber as an emerging pillar of the AI infrastructure thesis, that is confirmation rather than news.

2.5 What did not get answered

Vera Rubin ramp detail, standalone CPU strategy, China re-entry timing and gross-margin trajectory were the four items flagged in the day two note as the things to listen for. None was resolved on the verifiable public record.

The margin question is the live one. FQ2 gross margin already compressed on memory cost, Nvidia has notified customers of price increases above 15 per cent on Vera Rubin and Grace Blackwell for early 2027, and the Vera CPU's LPDDR5X allocation has reportedly been cut to 96GB per SOCAMM, 768GB per CPU, from a 1.5TB design intent. Memory is expected to be around 20 per cent of AI rack bill-of-materials by 2027. A company passing through cost increases above 15 per cent on a bill of materials where one input is heading to a fifth of the total is a company whose margin trajectory is an open question, not a settled one.

The financing loop: Oracle, CoreWeave, and who needs the market

Day two flagged customer-financed supply appearing three times in one week, all equity-linked. Day three extended it in a more traditional direction: prepayments, bring-your-own-hardware, at-the-market equity, and high-single-digit term loans. Take the two disclosures in order.

3.1 Oracle FQ1 FY27, reported after the close

Guidance. FY27 revenue at least 90 billion dollars, up 34 per cent in constant currency. Non-GAAP EPS 8.10 dollars, up 18 per cent. FY27 capex 90 to 95 billion dollars, with net cash capex not to exceed 70 billion after customer prepayments. FQ2 revenue up 30 to 34 per cent, cloud up 64 to 71 per cent, EPS 1.85 to 1.93 dollars. Twenty billion dollars of at-the-market equity issuance completed. Expansion part-financed through prepay arrangements and bring-your-own-hardware models.

A correction to the day two note. That note carried RPO of 638 billion into the print. The reported figure is 664 billion. But the number to watch was never the level. It was the sequential change: plus 26 billion quarter on quarter against plus 209 billion year on year. Backlog growth is decelerating hard off an extraordinary base. That is the RPO-conversion question, and it is the right one to be asking.

The structure is the story

Ninety to ninety-five billion of capex, against a 70 billion net cash capex ceiling that only holds if customers keep prepaying, funded partly by 20 billion of equity issued into a falling tape. Utilisation at 97.9 per cent and a 20 per cent renewal premium say the asset earns its keep. The question is not the asset. It is the funding stack behind it and the long end it is priced against.

3.2 CoreWeave

Every GPU we have could be sold to multiple different clients.

Michael Intrator, Chief Executive, CoreWeave, adding the company is struggling to meet demand every day

Newer contract economics are improving because customer pricing is rising faster than power and memory costs. That is the single most useful sentence of the day for the neocloud margin question, and it went almost entirely unreported.

Q2 backlog around 104 billion dollars, plus more than 25 billion of additional commitments in early Q3. 2026 capex guidance raised to 35 to 39 billion from 31 to 35 billion. Q2 capex around 9.4 billion. July price increases of 25 per cent across the line-up held. A recent term loan priced at 8.2 to 9.2 per cent. The stock fell 4.9 per cent to 94.94 dollars, reversing an 11.7 per cent gain the previous session.

The cleanest capital-structure lesson of the conference

Demand is not the variable. Pricing power is not the variable, because a 25 per cent list increase in July stuck. The variable is that capex guidance rose by more than revenue and operating income guidance, and the incremental dollar is funded at 8.2 to 9.2 per cent while the 30-year sits at 5.37 per cent. Equity gets paid last in that structure, so equity sold the good news.

3.3 The ranking that matters into year-end

Rank the day three names by dependence on the capital market, because in a rising cost-of-capital regime that ranking does more explanatory work than any demand forecast.

Pricing power protects the equity only of those who do not need the market. That sentence is the whole of day three compressed into one line, and it is the reason the capital-structure lens is not one input among several right now. It is the sorting mechanism.

The efficiency shock: DeepSeek V4.1-Flash versus the memory floor

On day two, the NAND industry told investors that prices are protected. SanDisk described eight partnerships covering about two-thirds of its FY28 bit supply with price floors that protect an 80 per cent gross margin. Kioxia said prices had already risen enough. The bull case for memory is that the bits needed per unit of AI work keep going up while supply is contracted out years in advance.

On day three, DeepSeek released a model that needs roughly a quarter of the memory per token its previous model needed. Memory shares fell. Bloomberg's headline the next morning called it a fresh reason for caution on memory stocks. That is the collision, and it arrived less than 24 hours after the floor thesis was laid out.

4.1 What was actually released

4.2 Reading it without flinching

The instinct is to file this as a repeat of the January 2025 DeepSeek selloff and move on. Do not. The mechanics are specific, and they cut two ways at once.

Bits per served token fell sharply. A KV cache of 890 bytes per token, against roughly four times that before, is a direct reduction in the DRAM and HBM footprint required to serve a given context at a given concurrency. This is precisely the variable the memory price-floor thesis is levered to.

Parameters per model nearly doubled. A 552 billion-parameter backbone plus 196 billion of conditional-memory modules is more capacity to hold, not less. Capacity is stored in memory too. The direction of the net effect depends on the serving-to-training mix, and nobody has that mix.

Token volume should rise. Cheaper serving raises agent concurrency; concurrency raises tokens; tokens raise compute. The Jevons argument is intact, and Nvidia's own framing of agentic demand as parabolic depends on exactly this mechanism.

So the net is compute-positive and memory-intensity-negative. Good for the compute layer, ambiguous to negative for the part of the memory thesis that assumes bits per unit of work only ever climb.

It resolves on a date we know

Micron reports FQ4 on 30 September, having guided to a record of roughly 50 billion dollars at the midpoint with an 86 per cent gross margin. If contract ASPs hold into that print, the eight-partnership floor structure is genuinely structural and the DeepSeek reaction was a headline trade in a jumpy tape. If contract pricing wobbles, the whole 80-per-cent-gross-margin-through-FY30 construction gets repriced, and so does the memory-driven share of the wafer-fab equipment cycle.

The honest position is that nobody knows yet. What can be done is to name the observable that settles it, which is contract ASPs and not spot, and to be clear that compute exposure benefits either way while memory-intensity exposure does not. Anyone claiming to know the answer before 30 September is guessing with confidence.

The rest of day three

5.1 SpaceX

The most thesis-relevant non-semiconductor session of the day, and the one with the weakest source quality. Treat every figure below as unconfirmed: the transcript summary carries stale-year artifacts typical of machine-generated write-ups.

A hosting deal described as adding 1.11 billion dollars a month from 1 December, roughly 13 billion dollars of incremental annual recurring revenue. Terrestrial compute deployment targeted at 5 to 10 gigawatts, up from a little over 2, monetised at 30 to 50 dollars per watt, with SpaceX claiming the high end of that range. Orbital compute framed as the most misunderstood opportunity: first orbital compute satellite launches, modified V3 buses with larger solar arrays, scaling through 2028.

Capital allocation in three parts: Starship as the enabling platform, terrestrial compute as the funding source, orbital compute as the option. Around 100 billion dollars of annual recurring revenue as the stated target. Starship Flight 13 delivered V3 demonstration units, relit Raptor and recovered after a precision splashdown. Flight 14 this month carries production V3 Starlinks. Starlink in 170-plus countries; direct-to-device V2 with full 5G targeted for the first half of 2028.

A private company just joined the power queue

If the 5 to 10 gigawatt figure is real, SpaceX becomes a competitor for exactly the land, power and shell capacity Nvidia says it is tracking, and a buyer of exactly the transformers, switchgear and thermal equipment the power and grid layer sells. The 30 to 50 dollars per watt monetisation needs reconciling against the Akamai figures from day two, 15 to 20 million dollars of revenue per megawatt for GPU and 25 to 30 for CPU, before either number belongs in a model.

5.2 Uber

Uber One above 50 million members, up 50 per cent year on year, roughly half of bookings, members spending about three times non-members and twice as likely to use multiple platforms. Access to more than 120,000 AV-ready vehicles; nine AV partners today, expected to reach 15 by year-end; no-vehicle-operator partnerships to be demonstrated next year; scaling expected 2028 to 2029. Top-line growth above 20 per cent for four consecutive quarters with simultaneous margin expansion. Grocery and retail delivery at 15 billion dollars, growing above 40 per cent. Sparse markets growing 1.5 times the core.

The read-through: Huang named physical AI and autonomous driving as a two-to-three-year growth leg, and Uber is the demand aggregator that has to absorb it, on the same 2028 to 2029 horizon. Two independent sessions converging on the same date range is worth more than either one alone.

5.3 Airbnb

AI is a wave, and we can surf the wave rather than get swallowed underneath.

Brian Chesky, Airbnb, who also compared AI to nuclear power and said he is not confident he is right

No investment content. It matters as sentiment. The safety conversation reached the main stage of a sell-side technology conference, one day after a high-profile researcher departure generated end-of-decade risk headlines. Narrative risk to the sector's multiple is no longer purely a rates story, and that is a new variable rather than a recurring one.

Where the framework is most exposed

Three days of conference content have sharpened rather than settled the picture. Four observations on where the analytical framework carries the most risk.

What to watch

7.2 Dated catalysts

7.3 Open questions carried forward

Data-integrity flags on this note

Conference coverage in the first 24 hours is assembled from machine-generated transcript summaries and second-hand reporting. The following items are known-weak. They are flagged so they are not laundered into a model or a downstream document by accident.

Glossary

ARR

Annual recurring revenue. Contracted revenue expressed as an annual run-rate.

ASP

Average selling price. Contract ASP is the negotiated long-term price; spot is the open-market price. They can diverge for quarters.

ATM offering

At-the-market equity issuance. Selling new shares into the open market over time rather than in one block. Oracle completed 20 billion dollars of this.

Bear steepener

Long-term yields rising faster than short-term yields. Compresses valuation multiples on assets whose cash flows sit far in the future.

BoM

Bill of materials. The component cost stack of a finished system.

CoWoS

Chip-on-wafer-on-substrate. TSMC's advanced packaging process; a named capacity constraint for AI accelerators.

D2D

Direct-to-device. Satellite connectivity straight to an ordinary handset, without ground equipment.

DRAM, HBM, NAND

Volatile working memory; stacked high-bandwidth memory used beside AI accelerators; non-volatile flash storage.

Engram modules

DeepSeek's term for sparsely accessed conditional-memory parameter blocks sitting outside the main backbone.

FCF

Free cash flow. Operating cash flow less capital expenditure. Negative for Oracle this quarter by design.

FOMC

Federal Open Market Committee. The Federal Reserve's rate-setting body. Meets 15 to 16 September 2026.

FP4 KV caching

Storing the attention key/value cache in four-bit floating point. Roughly quarters the memory footprint per token versus common alternatives.

Jevons paradox

When efficiency gains lower unit cost enough that total consumption rises. The core bull argument for compute after any model-efficiency release.

KV cache

Key-value cache. The memory an AI model holds per token of context while generating. The dominant memory cost in serving long-context and agentic workloads.

LPDDR5X, SOCAMM

Low-power DRAM standard; the memory module format used beside Nvidia's Grace and Vera CPUs.

MoE

Mixture of experts. An architecture where only a fraction of total parameters activate per token, hence 552bn total but 8 to 16bn active.

Neocloud

A GPU-first cloud operator outside the big hyperscalers. Named by Nvidia: CoreWeave, Nebius, Nscale, Lambda, Firmus.

OCI

Oracle Cloud Infrastructure. The 7.4 billion dollar, plus 121 per cent segment.

PPI, CPI

Producer price index, the prices firms receive; consumer price index, the prices households pay. PPI printed 10 September; CPI prints 11 September.

Prepayment, BYOH

Customer cash paid ahead of delivery; bring-your-own-hardware, where the customer supplies the accelerators. Both shift capex funding off the operator's balance sheet.

RPO

Remaining performance obligations. Contracted revenue not yet recognised, meaning the backlog. Oracle: 664 billion dollars.

TAM

Total addressable market. Huang's version: 3 to 4 trillion dollars of AI infrastructure spend by 2030.

Term premium

The extra yield demanded for holding long-dated bonds. Its expansion is what puts the 30-year at 5.37 per cent.

Vendor financing

A supplier funding its own customer's purchase, whether through equity warrants, prepayments or credit. The defining structure of the 1999 telecom buildout, and re-emerging here.

Vera Rubin, Grace Blackwell

Nvidia's next-generation and current-generation rack-scale platforms. Both carry announced price increases above 15 per cent for early 2027.

WACC

Weighted average cost of capital. The blended cost of a company's debt and equity funding.

WFE

Wafer fab equipment. The semiconductor capital-equipment spending cycle. Applied Materials moved its 2026 growth call from 20 per cent toward 40 per cent on day two.

Sources

TechCrunch, “Jensen Huang explains why Nvidia will grow an astounding 70% next year,” 10 Sep 2026 · Yahoo Finance / Barchart, “NVIDIA CEO Sees $4 Trillion AI Infrastructure Boom as Demand Outpaces Supply,” 10 Sep 2026 · Yahoo Finance, “Jensen Huang Maps Nvidia's Next Two Growth Engines,” 10 Sep 2026 · StreetSignal, “Jensen at Communicopia,” 10 Sep 2026, partially paywalled · Yahoo Finance, “Memory Prices Heat Finally Reaches Nvidia,” GF Securities / Jeff Pu on Vera CPU LPDDR5X · TrendForce, Nvidia Vera Rubin and Grace Blackwell price increases above 15% for early 2027, 24 Aug 2026 · Investing.com, “Oracle Q1 FY2027 slides: cloud infrastructure surges 121%, guides $90B+,” 10 Sep 2026 · Invezz, “CoreWeave cannot get enough Nvidia GPUs: why the stock still dropped 5%,” 10 Sep 2026 · Investing.com, “SpaceX at Goldman Sachs Communacopia + Technology Conference 2026: compute push,” 10 Sep 2026 · Yahoo Finance / Barchart, “Uber Technologies Targets Sparse Markets, AV Scale as Uber One Membership Tops 50 Million,” 10 Sep 2026 · CNBC, “Humans need to surf the wave of AI rather than get swallowed by it, Chesky says at Communacopia,” 10 Sep 2026, via syndication · VentureBeat, “DeepSeek-V4.1-Flash debuts with $0.003/1M off-peak cached-input rate,” 10 Sep 2026 · TechTimes, “DeepSeek V4.1-Flash Cuts Agent Memory Costs Fourfold With New Architecture,” 10 Sep 2026 · Bloomberg, “DeepSeek Hands Memory Stock Investors Fresh Reason for Caution,” 11 Sep 2026, headline only · TheStreet, “Stock Market Today (Sept. 10, 2026)” and “Fed rate-hike threat heats up as August inflation looms” · Investrade, “Market Review: September 10, 2026” · Seoul Economic Daily, “TSMC August Revenue Hits Record, Surging 53.3% From Year Earlier,” 10 Sep 2026.