$3.15bn Guided, the Margin Floor Broken (27 August 2026): Marvell Cleared the Q3 Bar at a 52% Incremental Margin — and Google's Warrant Starts Vesting Now

Marvell reported second-quarter fiscal 2027 revenue of $2,739.3m, a record, up 36.5% year on year and $39.0m above the mid-point of its own guidance. Non-GAAP earnings per share were $0.94 against a $0.93 guide. Both were, as expected, non-events — the quarter had been pre-announced in May and consensus had converged on it. The two numbers that carried information both sit in the third-quarter outlook, and they point in opposite directions. Revenue is guided to $3.150bn, roughly $120m or 4% above the $3.03bn consensus bar. Non-GAAP gross margin is guided to 57.5–58.5%, a mid-point of 58.0%, breaking below the 58.25% floor that had contained it. Put the two together and the guide implies an incremental non-GAAP gross margin of about 52% on the added revenue, against the 59.2% Marvell just delivered sequentially. The acceleration is being bought. And the reason is now documented rather than inferred: a Google custom-silicon agreement whose 240 performance tranches begin vesting in this exact quarter.

What actually happened

The printed quarter did what a pre-announced quarter does. Revenue of $2,739.3m beat both the company's $2.700bn mid-point and the $2.71bn consensus; non-GAAP EPS of $0.94 beat the matched $0.93 by a cent. GAAP gross margin landed at 53.1%, the top of its guided band. The interest is in the second column.

All figures are from Marvell's second-quarter fiscal 2027 release and its accompanying supplemental tables. The revenue acceleration management promised in May is arriving on schedule: 27.6% year-on-year growth in the April quarter, 36.5% in July, and 51.8% guided for October.

The segment split is where the shape shows. Data centre revenue was $2,171.5m, up 45.7% year on year, and reached 79% of the company against 74% a year earlier. Communications and other was $567.8m — and it fell 3.0% sequentially, from $585.1m. So of the $321.5m the company added sequentially, the data centre segment contributed $338.8m and the rest of the business gave back $17.3m. There is one engine now.

The bar cleared, the floor broken

The third-quarter guide is genuinely two events wearing one headline, and reading only the first one gets the quarter wrong.

Non-GAAP gross margin ran 59.4% in the July 2025 quarter, 58.9% in April 2026 and 58.9% again in July 2026. The Q2 band bottomed at 58.25%. The new band tops out at 58.5% — barely above where the old one ended — and mid-points at 58.0%.

The 52% that the guide implies

The useful operation is to take the guide apart into base and increment. Non-GAAP gross profit in the quarter just reported was $1,614.0m on $2,739.3m of revenue. Apply the guided margin band to the guided revenue and you get the gross profit the company expects; the difference between the two, over the difference in revenue, is the incremental margin on everything new.

That is the finding. Marvell just delivered a sequential incremental gross margin of 59.2% — above its own average, not below it — and then guided the next quarter to somewhere between 48% and 56%, mid-point 51.9%. The compression is not in the quarter that happened. It is entirely in the quarter being promised.

The last step matters and is usually lost. Non-GAAP operating expenses are guided to roughly $655m against $610.8m actual — up 7.2% while revenue is guided up 15.0%. Run it through and implied non-GAAP operating income is about $1,172m, a 37.2% operating margin against the 36.6% just delivered. So the gross margin falls and the operating margin rises in the same guide. Both are true; they are measuring different things. Gross margin is telling you about the price of the product, and operating margin is telling you about the scale of the company.

The 240 tranches, and the quarter they start

The preview version of this note treated the mix shift as a structural inference. It no longer has to be. On 29 July 2026 Marvell and Google LLC entered a commercial agreement for Marvell to develop custom semiconductor products for Google, covering programmes that attach to the TPU ecosystem: AI inference accelerators, storage controllers, network interface controllers, memory interface controllers and near-memory compute. In connection with it, on 18 August 2026 Marvell issued Google a warrant over 58,970,907 shares at an exercise price of $206.58, disclosed in a Form 8-K filed on 19 August.

The vesting mechanics are the part worth reading closely.

Three pieces of arithmetic follow. First, 240 tranches at $500m each means full vesting requires $120bn of Google custom-product purchases, running from the third quarter of fiscal 2027 to the end of fiscal 2033 — about 26 quarters, or an average of roughly $4.6bn per quarter. Marvell's entire company revenue is guided to $3.15bn this quarter. Full vesting is a tail, not a base case, and the tranche structure is what makes that legible.

Second, the dilution at full vest is 58,970,907 shares against 921.2m diluted, or 6.40%. Per tranche it is 0.026% of the diluted count for each $500m of purchases. At $206.58 the notional exercise proceeds across the whole warrant are about $12.18bn.

Third, and this is the connection to everything above: the performance tranches accrue on purchases from the third quarter of fiscal 2027. That is the quarter Marvell just guided, at the margin that just broke the floor. The customer whose volume arrives now is also the customer earning equity on it.

Why GAAP EPS missed a guide the revenue beat

GAAP diluted EPS was $0.33 against a $0.37 guided mid-point — a miss on the company's own number, in a quarter where revenue beat. The gap to the non-GAAP $0.94 was $557.9m, or $0.61 per share, against the $0.56 that had been guided. Two lines explain it.

Stock compensation rose to $326.2m from $153.6m a year earlier — 11.9% of revenue against 7.7%. Within it, the research and development line more than doubled year on year to $219.2m and selling and administrative went from $36.1m to $91.1m. It is now the single largest non-GAAP add-back, larger than intangible amortisation, and larger than the entire GAAP net income of $308.0m.

The second line is a correction to the preview. That note described the contingent-consideration mark as episodic and specific to the first quarter, and expected it not to repeat in size. It repeated: another $101.9m, taking the six-month total to $433.7m. The liability is a deal obligation being revalued upward, and two consecutive quarters of it is a pattern rather than an event.

Cash conversion is worth one line. Operating cash flow was $605.5m against non-GAAP net income of $865.9m, or 70%, and it fell sequentially from $638.8m even as revenue rose $321.5m — receivables absorbed $346.6m as the balance built to $2,218.4m. Free cash flow after $131.2m of capital and licence spending was about $474.3m. The company returned $200.0m in buybacks and $53.9m in dividends.

The other half of the market got benchmarked

Marvell's release priced the custom-silicon trade from the supplier's side. Two days earlier, the same trade was measured from the buyer's side.

On 25 August 2026 OpenAI published the first benchmarking results for Jalapeño, the inference chip it announced in June and is developing with Broadcom, saying the part had industry-leading speed and efficiency. Deployment inside OpenAI's own compute infrastructure is due to begin by the end of the year, and the company said generations two and three were already in development. The research firm SemiAnalysis, which went to OpenAI's labs to run the part, published its findings the same day under the title OpenAI Jalapeño: Better Than Nvidia Blackwell, reporting that the chip beat Blackwell systems on performance per watt in nearly all tested scenarios. The account of both is CNBC's, published 26 August.

That result is the demand-side explanation of Marvell's margin line. A buyer commissions a custom part because the efficiency gain accrues to the buyer rather than to the merchant supplier it displaces, and the saving is larger than the chip. Alexander Harrowell, senior principal analyst at Omdia, framed it as a facilities question rather than a silicon one: in a large-scale deployment, he told CNBC, the efficiency "would save power, cooling, and power distribution infrastructure, and contribute a lot to their unit economics". Those savings land on the buyer's cost base. The designer that made them possible books the socket at the 52% incremental margin the section above worked out. That is not two stories. It is one transfer, seen from each end.

The scale question is what makes this a sector datapoint rather than a company one. Custom accelerator programmes are no longer a fringe of the buildout, and the named ones now cover most of the buyers who matter.

Harrowell put a size on where that leads: "Omdia expects custom ASIC chips like Jalapeño to exceed GPUs in volume by 2028, although revenue will take much longer as GPUs are considerably more expensive." He called it the biggest competitive threat to Nvidia, on the grounds that about half of capital expenditure on AI infrastructure comes from hyperscale cloud providers that either run a custom chip programme or could plausibly start one. Note what that forecast does and does not say. Volume crossing over in 2028 while revenue lags is precisely the mix shift Marvell just guided: more units, priced against a known cost structure, at a lower percentage margin. The forecast and the guide describe the same arithmetic at different points on the curve.

Two limits are worth holding onto, because they are the reason this is a read-across and not a verdict. The first is workload. Fion Chiu, an analyst at TrendForce, told CNBC that a custom chip could reduce OpenAI's reliance on Nvidia over time for inference, but that for compute-intensive work such as large-scale model training and frontier workloads, "we believe Nvidia GPUs will remain important given their broad programmability, performance, software ecosystem, and ability to handle a wide range of workloads". Custom silicon is taking the repetitive half of the job, not the exploratory half. The second is the software estate. Adrien Sanchez, technology analyst at Yole Group, said Jalapeño shows that a hyperscaler-designed chip can now match or beat Nvidia's Blackwell-class GPUs on inference efficiency, while noting that Nvidia still owns the vast majority of AI compute and has ecosystem lock-in through CUDA. His characterisation of the risk is narrow and specific: a "threat to Nvidia's inference margins, which is the field growing the most at the moment." Margins in one segment, not the franchise.

Which brings the thread back to a date. Broadcom announces third-quarter fiscal 2026 results after the close on Wednesday 2 September 2026, having guided AI semiconductor revenue to $16.0bn for the period. Marvell has now disclosed what the custom template costs at its own scale: roughly 52% on the increment, plus equity to the customer. Broadcom is the same template at several times the volume, reporting five business days later, with its flagship customer programme freshly benchmarked in public. Whether the percentage margin behaves the same way at that scale is the open question, and 2 September is when it stops being open.

The channel to what you actually trade

Marvell is a Nasdaq-100 constituent but not a weight that moves NAS100 or US500 on its own. What travels is the read-across, and this release sharpens it in a specific way. Until now the custom-silicon story could be told as pure volume growth. This guide prices it: roughly 52% incremental gross margin at the mid-point, plus equity to the customer, in exchange for 15% sequential revenue growth. That is a repeatable template, and Broadcom — the other half of the custom accelerator market, reporting on 2 September against a $16.0bn AI semiconductor guide — is the next test of whether the same terms apply at five times the scale.

The read-across runs in both directions with the merchant side of the market. The Nvidia quarter reported on 26 August guided its own gross margin down for reasons of its own, and the wider supply chain that populates the same racks takes its cue from the same demand, a dynamic traced in the note on mature-node chip pricing.

For currencies this is not an FX event, and pretending otherwise would be dishonest. The dollar's week belongs to the rates picture and the Jackson Hole symposium, which is why the dollar page and the five factors behind it are the right place to read that leg. The one real crossover is risk sentiment, which reaches the pro-cyclical currencies from one side and the yen and the franc from the other — a second-order channel, not a first-order one.

What would change the picture

Whether the October quarter lands in the top or bottom half of the 57.5–58.5% band, since that is the difference between a 55.7% and a 48.0% incremental margin and therefore between a mix shift that is grinding and one that is accelerating. Whether the quarterly filing discloses the warrant as consideration payable to a customer, which would put part of the cost inside revenue rather than beside it. Whether communications and other resumes growth or keeps shrinking, because a single-engine company and a two-engine company carry different concentration risk under the same revenue line. Whether stock compensation stabilises near 12% of revenue or keeps climbing, since it is now the largest wedge between the two earnings figures. Whether Broadcom's 2 September report shows the same incremental-margin behaviour at its scale, since that is the difference between a company-specific trade and the price of the whole custom-silicon template. And what the raised fiscal 2027 and 2028 revenue outlooks actually are: the first half delivered $5,157.1m, and adding the $3.150bn guide puts three quarters at $8,307.1m, which means the near-$11.5bn frame from May would need only about $3.19bn in the January quarter — a 1.4% sequential step after four quarters of accelerating growth. That arithmetic is why the outlook had to go up.

The demand was never the question in this print. The price of supplying it was, and the guide answered.

More on the method behind these notes is on the about page.

Educational macro context only — not investment advice.