The Rest of 2026: The AI Margin Transfer, and Why the Fed Only Matters Below the Index

Sunday, August 30, 2026 · All prices are Friday, August 28, 2026 closing prices (US markets closed for the weekend) · @dailyanalysts

1. What happened, and why Friday was the tell

Kevin Warsh gave his first Jackson Hole keynote as Fed chairman on Friday and told the market the Fed may "have work to do." September rate-hike odds went from 35.4% to 57.5% on the CME FedWatch tool in a single session (CNBC). The 2-year yield jumped 12.8bp to 4.36%. The 10-year closed at 4.73%, the 30-year at 5.213% — near multi-decade highs.

And the S&P 500 closed at 7,715.00, down 0.10%.

The index absorbed a near-doubling of hike probability with a shrug. That is the entire story of the rest of this year in one data point: the Fed no longer moves the index. It moves what is underneath the index. Look at where Friday's damage actually landed:

Four of the five largest companies in the world went up on a hawkish Fed day. Every company that sells shovels to them went down 2–10%. Gold fell 3.24% and silver 4.38% on the same news. That is a coherent, single-factor day, and the factor is not "risk appetite." It is who funds the AI buildout out of operating cash flow, and who has to finance it.

2. The core argument: NVIDIA just handed $16bn a year to the memory makers, and told you it would

NVIDIA's Q2 FY2027 release (August 26) was, on the surface, the best print in the history of the semiconductor industry. Revenue $96.221bn, up 106% year over year. Data Center $89.0bn, up 117%. Gross margin 75.0%. And a first-ever forward-year guide: FY2028 revenue growth of approximately 70%, versus a Street modelling roughly 44% — implying ~$690–700bn against consensus near $570bn, with management stating the number is supply-constrained and true demand is closer to 100%.

Every headline led with that. Almost nobody led with the guided margin path, which management reset in one motion on the call:

Do the arithmetic, because it is the whole trade. Peak-to-trough is 350 basis points. Q3 revenue is guided to $108.0bn and the sequential trend puts Q4 near $118bn. At an average of roughly $120bn a quarter through the trough:

That is the single most consequential number in the print, and it is the organising fact for the next four quarters. Memory is sold out. Samsung and SK Hynix have reportedly sold DRAM and HBM through 2027; SK Hynix's CEO has said the worst of the shortage is 2027, not 2026. CSIS documents the downstream damage: memory is 15–18% of a typical PC's bill of materials, PC OEMs have pushed 15–20% price increases, and TrendForce cut its 2026 notebook forecast from +1.7% growth to a 2.4% decline.

Micron's numbers show where the margin went: revenue +167% y/y, gross margin 72.57%, ROE 70.55%, a $1.087 trillion market cap. Micron now earns a gross margin within 250bp of NVIDIA's, in a business that historically earned 30%.

Consequence one: for the remainder of 2026, the correct expression of "AI capex is accelerating" is not long the accelerator layer. The accelerator layer is the one absorbing the input cost inflation. It is long the physically capacity-constrained layers — lithography and memory — that set the price.

Consequence two: NVIDIA's Q4 report (late November) contains a guided-down margin print that is already public. Anyone buying NVDA here is buying into a known negative catalyst with the offset (price increases) landing a quarter later. That is a timing problem, not a thesis problem, and it explains a good deal of Friday's 4.58%.

3. The thing consensus has not modelled: NVIDIA is increasingly financing its own demand

My readers know the technology better than I do. So let me stay where the edge actually is — the balance sheet and the cash flow statement, which almost nobody read.

Now the cash flow statement, which is where it becomes concrete:

Revenue rose 18% sequentially. Free cash flow fell 56% sequentially. The bridge is $22.3bn of receivables build in a single quarter plus $5.8bn of inventory.

Convert it to days sales outstanding, which is the honest way to see it. Q4 FY26 revenue was $68.1bn against $38.47bn of receivables — 51.4 days. Q2 FY27 revenue was $96.22bn against $63.06bn — 59.6 days. Collection has slowed by 8.2 days, which at the current run rate is roughly $8.7bn of revenue that has been recognised but not collected relative to the old cadence.

Why does that matter now and not two years ago? Because the customer has changed. NVIDIA disclosed the Data Center split for the first time this quarter:

The bull case for this disclosure is that customer concentration is resolving. That is true and it is genuinely good news. The part nobody has priced is the corollary: the fastest-growing 45% of NVIDIA's data centre revenue is being sold to counterparties with a materially worse credit profile than Microsoft and Amazon. Neoclouds, sovereigns and mid-tier enterprises do not pay in 30 days the way a hyperscaler does. That is exactly what 8.2 days of DSO expansion looks like when it shows up in the accounts.

And NVIDIA is on both sides of the transaction. In the same quarter it:

• Raised $24.9bn of new debt (long-term debt went from $7.5bn to $32.4bn)
• Returned $25.8bn to shareholders ($19.7bn buybacks + $6.0bn dividends)
• Bought $15.8bn of equity securities ($42.4bn in the first half), taking combined marketable + non-marketable equity stakes from $35.1bn to $93.9bn in six months
• Announced financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilise over $500bn of third-party capital for AI infrastructure
• Booked $7.77bn of gains on equity securities into GAAP other income — $23.7bn in the first half, or 20.1% of first-half GAAP net income, which is why GAAP EPS ($4.85 H1) exceeds non-GAAP ($4.09)

4. That event is the Anthropic IPO, and it is the most important thing on the calendar

Anthropic filed confidentially in June 2026 and investors expect an October Nasdaq listing at a $2 trillion-plus valuation, per reporting collated by R&D World from the Financial Times and CNBC. Some backers model $3 trillion. Run the numbers:

The load-bearing assumption in the entire $2T ask is that the 5.1% Q2 operating margin is structural. It probably is not. That margin was earned during a window when enterprises were spending on token volume with few cost controls — Uber and Microsoft have since imposed caps. Meanwhile Ramp's August index shows Anthropic's paid business adoption growth slowing, and use of model-serving platforms offering cheaper open-weight and Chinese models rising from 4.5% to 6.1% of AI-spending businesses between January and July.

Against that, the commitments are locked: more than $130bn to AWS and Azure compute alone, plus an undisclosed multi-gigawatt Google/Broadcom deal, plus $1.25bn a month to SpaceX. Anthropic raised $95bn in 2026 — roughly the combined annual R&D budgets of Alphabet and Microsoft.

Here is the connection nobody is drawing. The Anthropic listing is the mark-to-market for the entire private AI capital stack. It prices at $2T and holds → every private AI mark is validated, the neocloud and sovereign funding channel stays open, NVIDIA's ACI&E segment keeps compounding at 138%, and the receivables build above is an accounting curiosity. It breaks issue → the funding channel for the fastest-growing 45% of NVIDIA's data centre revenue tightens, and that $8.7bn of stretched receivables stops being a curiosity.

You do not need to own a share of Anthropic for this to be the most important trade input of your fourth quarter.

5. The upstream: ASML is the best risk-adjusted AI position available, and the screens are lying about why

Pull up ASML on any provider and you will see revenue growth of 9.84% year over year against a P/E in the high 50s. That is a stale trailing number and it is the reason the stock is mispriced.

Go to the company's own July 15 release. ASML entered 2026 guiding €34–39bn of net sales, which it described as 4–19% growth — implying a 2025 base of about €32.7bn. It has now raised full-year 2026 guidance to €43–45bn. At the €44bn midpoint that is +34.6% growth, not 9.8%. A mid-year raise of roughly €8bn is not a demand forecast. It is booked backlog.

And the CEO statement contains the sentence that changes the business model:

EUV capacity through 2027 is already fully booked. Which means ASML's 2027 revenue is no longer set by customer demand. It is set by ASML's own throughput. That is the conversion of a violently cyclical business into a capacity-scheduled one, and it is what a 40× multiple is actually paying for — not the 9.8% trailing growth number on the screen.

The model, and where it could be wrong

2026 net sales: Q1 €8.767bn + Q2 €9.326bn = €18.093bn. Q3 guided €11.0–12.0bn. To reach the €43–45bn full-year range, Q4 must land between €13.4bn and €15.4bn — a 22–28% sequential step above a Q3 that is itself a record. That is aggressive, and it is the one thing to check in mid-October.

My 2026 EPS build: H1 actual €14.74. Q3 at €11.5bn and 56% GM, less ~€1.6bn opex, gives ~€10.45. Q4 at €14.4bn and 57% GM gives ~€14.05. FY2026 EPS ≈ €39 on roughly 385m shares — about €15bn of net income against the ~$600bn market cap, so roughly 40×.

2027: strip out Installed Base Management sales (€5.25bn in H1, annualising near €11bn) and 2026 system revenue is about €33bn. A +30% capacity add that is fully absorbed takes system revenue toward €41–43bn; IBM grows with the installed base to ~€13bn. 2027 net sales ≈ €54–56bn, EPS ≈ €52.

Load-bearing assumption: that the +30% 2027 capacity add is fully absorbed. If only half of it is (+15%), 2027 revenue lands near €50bn and EPS near €45. At 36× that is roughly $1,780 — barely 5% above the current price. At 30× it is about $1,480, which is why the invalidation sits at $1,450. If you are wrong about absorption, you lose the trade, not just the upside. Watch the Q3 print in October: a low-end €11.0bn with soft Q4 guidance is the first crack.

6. Marvell: the lesson is about duration, and I owe readers an update

Marvell beat and raised, and fell 10.28% to $216.62. Q2 FY27 revenue $2.739bn (+36.5% y/y), non-GAAP EPS $0.94, data centre 79% of revenue growing 46%. Q3 guided to $3.15bn ±5% — roughly 50% year-over-year growth. FY27 outlook raised to ~$12bn, FY28 raised to ~$18bn. Non-GAAP operating margin expanded to 36.6%.

Fifty per cent growth got sold. Two reasons, and the second is the one that matters.

First, the stock was up 184.5% year to date and 47.8% in the four weeks into the print. Across the last seven reports, six were beats and the average day-of move was −2.09%. Clearing consensus and clearing an owned price are separate tests.

Second — and this is the real content — CEO Matt Murphy said the Google revenue for next year is already baked into guidance, and the bigger payoff comes the year after. Reuters reported the custom-chip deal will not become a major revenue engine until fiscal 2029. Marvell's own target is over $10bn of custom revenue in FY2029. So the $12.2bn Google warrant — up to ~59m shares at $206.58, tied to purchase targets through fiscal 2033 — was priced into the stock in August and the economics arrive in 2029.

Accounting for my open call

On August 28 I opened EQ-MRVL-1: contrarian long, entry $210–225, target $270, invalidation "weekly close below $195, or any disclosed cut to Google TPU volume or timeline." MRVL closed Friday at $216.62 — inside the entry zone, so the position is on.

Honest reading of my own invalidation clause: the FY29 framing is a deferral in emphasis, not a disclosed cut — no programme or volume was reduced, and both FY27 and FY28 guides were raised. So the invalidation did not fire. But it weakened the thesis, and I am tightening rather than pretending otherwise: target cut from $270 to $265, stop raised from $195 to $198, and a hard event check at the October 6 Investor Day. If Marvell does not attach firm FY28 revenue to named custom programmes that day, I am out regardless of price.

7. Broadcom, September 2: the cleanest binary of the quarter

AVGO closed at $368.79, roughly 25% below its June 3 high of $495, and reports fiscal Q3 after the close on Wednesday, September 2. One line decides it.

In June, Broadcom beat on earnings and fell 12.59% — because management held its fiscal 2027 AI target at "in excess of $100 billion" rather than raising it. In a stock priced for acceleration, a flat guide read as a stumble. Since then Marvell walked into the Google account (AVGO fell 5% on August 19) and Macquarie has modelled Broadcom's share of Google custom-chip spend falling from ~95% toward 65% by 2028.

The setup: Q3 AI semiconductor revenue is guided to $16bn, up over 200% y/y, and Q2 bookings topped $30bn against $10.8bn shipped — a 2.8× book-to-bill. NVIDIA has just told the market the total pie grows ~70% next year, not 44%. A bigger pie can absorb a smaller Broadcom slice. That is precisely the condition under which the 25% drawdown is wrong.

The offsetting risk is real and I will not soft-pedal it: Broadcom is reportedly in talks to help arrange $60–100bn of debt via an SPV to finance AI chips for Anthropic and others, guaranteeing a portion. Its credit-default swaps widened about 28bp on August 24. That is the same "vendor becomes the lender" structure flagged in Section 3, on a company already carrying customer concentration.

8. Rates, debt and the Fed: what the numbers actually oblige you to do

Core PCE for July came in at 3.3% annually, in line with forecasts; headline PCE was 3.7%, and CPI is running at 3.4% (CNBC). Warsh recommitted to the 2% PCE target as a "firm, fixed target" and noted that 54% of PCE components are running above 3% annualised. He also stated flatly that "short-term interest rates are the predominant tool," and that the Fed's AI task force has "no bearing on decisions we make in the current policy conjuncture" — closing off two of the dovish arguments he was appointed on.

The fiscal backdrop: total US debt has passed $40 trillion, the 2026 deficit is on track to exceed $2 trillion, and 2025 saw roughly $4.8trn of issuance. Treasury Secretary Bessent has proposed at least doubling debt buybacks and floated using the $935bn general account. His own mentor, Stanley Druckenmiller, publicly told him to stop: "If the 30-year must trade at 5.5% to clear, that isn't a crisis. It is an invoice." (CNBC)

Now the number that should govern your allocation. The S&P 500 closed at 7,715. On Goldman's May 2026 estimate of $340 of 2026 EPS, that is 22.7×, an earnings yield of 4.41%. The 10-year Treasury yields 4.73%.

The asymmetry most people have backwards. With a September hike 57.5% priced, a hike is now the expected outcome. The 42.5% branch — a hold — is the one that is not in the price. Given that the July employment report showed the US actually lost jobs, the August jobs print on Friday, September 4 is a live two-way catalyst, and a soft number would deliver more to high-duration assets than a hike would take away. Do not be positioned as though a hike is the risk. The hike is the base case.

Three asset classes, three regime-changing levels

9. The unpriced tail: chip tariff Phase Two

Presidential Proclamation 11002, signed January 14, 2026, imposed a 25% Section 232 tariff on advanced semiconductors — but carved out data centres, R&D, startups and consumer devices. Politico reported on August 27, from eight sources, that the administration is weighing a Phase Two that would do two things at once: extend duties downstream to laptops, consoles and data-centre servers, and eliminate the Phase One exemptions, including the data-centre carve-out (summary and sourcing).

The proposed mechanism — Commerce Secretary Lutnick's preference — ties duty-free import volume to committed US manufacturing investment. The structural problem is arithmetic: TSMC has committed $265bn to Arizona, the largest foreign direct investment in US history, and even at full build-out projects roughly 30% of its advanced capacity in the US, against Taiwan's current 90%-plus share of leading-edge output. A quota keyed to domestic capacity sets a ceiling far below what hyperscalers are consuming.

Quantify the mechanism, because that is the point. If the data-centre exemption is removed, a 25% duty lands directly on imported accelerators. Applied to NVIDIA's guided $108bn quarter, that is a ~$27bn per-quarter cost that must be absorbed somewhere in the chain — by the hyperscaler's capex budget, by NVIDIA's already-compressing gross margin, or by unit volumes. There is no fourth option, and no domestic supply alternative inside five years. Nothing has been announced and the framework may change substantially. But this is the largest genuinely unpriced risk to the AI complex for the rest of 2026, and it is not in a single sell-side model I can find.

10. Crypto: do not chase this

Bitcoin is at $78,174, Ethereum at $2,456.93, total crypto market cap $2.634trn with BTC dominance at 59.47%. The 14-day path is the story: BTC traded at $63,015 on August 16 and peaked at $80,698 on August 25 — a +28% two-week rally — before fading to current levels.

Three facts that sit badly together:

• Fear & Greed reads 69 (Greed), near the top of its own seven-day range of 65–74.
• US spot Bitcoin ETFs posted $201.8m of net outflows on Friday, ending a nine-day inflow streak, with total fund assets slipping back below $100bn (Cointelegraph).
• BTC fell roughly 2.4% during Friday's US session (from $79,438 at 14:00 UTC to $77,532 at 19:00 UTC), alongside gold −3.24% and silver −4.38%.

That last point deserves emphasis, because it kills a popular narrative. The "debasement trade" — long gold and bitcoin against fiscal deterioration — is supposed to be indifferent to the Fed. On Friday, with US debt above $40trn and the 30-year at 5.213%, gold, silver and bitcoin all fell together on a hawkish central banker. The debasement complex is rate-sensitive after all. If the Fed hikes on September 16 and the dollar firms, all three have further to fall.

My view, marked as opinion: a 28% two-week rally into greed readings, with ETF flows turning negative on the same day the Fed turned hawkish, is a place to trim, not to add. No new long above $80,700 — the August 25 high. A failure to reclaim $80,700 within two weeks while the Fed is hiking is distribution, and the downside re-test is the $70–72k shelf from the August 20 breakout. This is a WATCH, not a short: I have no edge calling the top of a momentum move and I will not pretend otherwise.

11. The trade table

Carried-forward calls, marked to Friday's close

• EQ-IWM-1 (trim/short IWM, entry $293–297, target $282, invalidation weekly close above $302) — IWM closed $295.75, −1.35%, inside the entry zone and the worst-performing major index on a hawkish day. Live and working.
• PC-HIKE-1 (long OBDC, entry $10.90–11.40, target $13.00) — OBDC closed $11.30, inside the zone. Live. Disclosed conflict, as required: EQ-IWM-1's $282 target runs through the $290 IWM level that is a warning condition for the floating-rate credit complex. If the IWM short works, it triggers a risk flag on OBDC on the way. Coherent as a pair, but readers following both must know it.
• PC-HIKE-4 watch levels — none triggered. 2Y at 4.36% (trigger 4.50%); IWM $295.75 (trigger $290); HYG $79.74 (trigger $78.00); BIZD not checked this session.
• CRM-CRWD-OKTA-TRIM — I am closing this as a partial miss and readers deserve the accounting. The trim into August 28 was directionally right, but the stated re-entry zones (CRM $215–225, CRWD $200–210) were never reached. Instead Salesforce surged ~22% and CrowdStrike ~20% on Thursday, August 27 on strong prints, and both finished the week as leaders. The call left a 20% move on the table. The re-entry zones are abandoned and the call is removed. The lesson, taken honestly: I mapped "high duration" onto "will fall on hawkish news" and the fundamentals overrode the rates factor.
• EQ-VIXHEDGE-1 — I could not obtain a verified VIX quote this session and I will not cite a level I have not confirmed. The thesis (cheap index protection into the September 4 jobs print and the September 16 FOMC) stands; I will re-mark the level when the data handler returns it.
• EQ-PYPL-WATCH — no new information. Remains a watch.

12. Scenarios for the rest of 2026 (September 1 – December 31)

13. What would prove me wrong

My thesis is that margin and leadership transfer upstream (to capacity-constrained litho and memory) and to cash-funded platforms, away from the accelerator layer and the debt-financed middle. Four specific, measurable things would break it:

1. NVIDIA's Q3 print (late November) shows gross margin above 74.5% and operating cash flow above 70% of net income. That would mean the memory squeeze and the receivables build were both single-quarter artifacts and the accelerator layer keeps its margin. This is the cleanest falsification and it has a date.
2. ASML guides 2027 low-NA EUV shipments below ~80 systems (against the ~85 implied by the +30% plan), or misses the €13.4bn Q4 required to reach the low end of its own FY26 range. Either directly breaks the "supply-scheduled" argument that the whole ASML call rests on.
3. DRAM contract pricing turns negative in any quarter before mid-2027. The margin transfer only works while memory is the bottleneck. One quarter of falling contract prices and the trade is over.
4. SMH makes a new closing high above $573 while MSFT, GOOGL, META and AMZN underperform over the same two weeks. That would say Friday's rotation was noise and I read a single session as a regime.

14. What to do

1. Make ASML the core AI position, not NVIDIA. Entry $1,600–1,720, target $2,050, out below $1,450. You are buying a business whose 2027 output is already sold and whose growth rate is set by its own capacity plan, at roughly 40× a 2026 that is growing 35% — not the 9.8% the screens show.
2. Hold NVIDIA; stop adding above $230. The Q4 margin trough is a known, guided, negative catalyst in late November and the price-increase offset lands a quarter later. Nothing here is a short. It is a timing problem, and you already own it.
3. Rotate the AI dollar from the financed middle to the funded top. Prefer Microsoft, Alphabet, Amazon and Meta — which pay for compute out of operating cash flow — over neoclouds and 60×-forward custom silicon that depend on capital markets staying open.
4. Wait for Broadcom's September 2 guide rather than guessing it. Buy above $385 on a raise. If the FY27 AI number is held flat again, do nothing — the June precedent is a 12.59% single-day drop.
5. Put real money in 2-year Treasuries at 4.36%. The S&P's forward earnings yield is 4.41%. You are being asked to take equity risk for five basis points. This is the diversification that actually diversifies, and it is uncorrelated to every other line on this page.
6. Do not chase bitcoin above $80,700. Twenty-eight per cent in two weeks, greed at 69, ETF flows negative on the day the Fed turned hawkish.
7. Put the Anthropic IPO on your calendar as a macro event, not a tech event. Expected October. It is the mark-to-market for the private capital that funds 45% of NVIDIA's data centre revenue.