Three AI Hardware Beats and Three Stocks Down as the Market Prices Margins. Cisco, Coherent and Cerebras All Fall.

Thursday, August 13, 2026, written at 9:20am Eastern, fifty minutes after the PPI release and ten minutes before the bell, so the marks below are live and several will move within the hour. Three companies that sit at the physical centre of the AI buildout reported after yesterday’s close. All three beat on revenue. All three are lower this morning. That is the story, and it is a different story from the one this note told yesterday.

Top Call: the Market Repriced the Margin Line, Not the Order Book

Cisco reported fiscal Q4 revenue of $17.3 billion, up 18%, with non-GAAP EPS of $1.22, up 23%, both above the guidance range, and disclosed $9.3 billion of hyperscaler AI infrastructure orders for the fiscal year, roughly 4.5 times the prior year, with about $4 billion of that booked in the June quarter alone. It guided fiscal 2027 revenue to $72.2 billion to $73.4 billion, up 15% at the midpoint, including $7.5 billion of AI infrastructure revenue against roughly $4 billion delivered in fiscal 2026. The stock is down more than 6%.

Coherent reported record fiscal Q4 revenue of $2.05 billion, up 34%, with datacenter and communications revenue up 59% and now 79% of total sales, non-GAAP EPS of $1.74, up 74%, and guided the September quarter to $2.2 billion to $2.4 billion against $2.13 billion of consensus with earnings 10% above the street at the midpoint. The stock is down about 5%. Cerebras posted core revenue of $210 million against roughly $194 million expected, with core cloud revenue up 287%, and raised full-year guidance. The stock is down about 15%.

The common line in all three is gross margin. Cisco’s non-GAAP gross margin fell to 66.3% from a year ago and management guided lower on a heavier hardware mix. Cerebras guided gross margin to 36% to 38% from 47% in the prior quarter, nearly ten points of sequential compression. Coherent’s margin is actually expanding, at 40.2% non-GAAP against 39% to 41% guided, and it still got sold, which tells you the reaction is about the multiple the market is willing to attach to AI hardware revenue rather than about any one company’s cost line.

Our take: this is a genuine regime shift and it is healthier than it looks on the tape. Yesterday the market stopped asking whether the buildout could be funded because Super Micro and CoreWeave produced order books too large to argue with. Today it has started asking what that revenue earns, and the answer arriving from three different layers of the stack, networking, optics and silicon, is that AI mix is dilutive to gross margin in almost every business that touches it. Stay long the complex and shift the screen from bookings growth to incremental gross margin on those bookings. The companies that can hold margin while the mix shifts are about to separate violently from the ones that cannot.

The Tape

Wednesday’s cash session was constructive without being decisive, with the S&P up 0.3% to 7,748.50 in its first gain since the record set last Friday, the Nasdaq Composite up roughly 0.5% to 26,588.49 and the Dow flat at 53,770.27. This morning the futures picture is unusually informative because of what is not leading. Dow futures are the strongest of the four at +0.26% and Nasdaq 100 futures the weakest at +0.06%, an exact inversion of yesterday’s leadership, and the reason is one stock. Cisco is a Dow component and a heavyweight in the Nasdaq 100, and the fact that the broad tape is absorbing a 6% drop in a company that size without breaking stride says the rotation is happening inside the complex rather than out of it.

Cisco: the Best AI Order Number Any Networking Company Has Reported, and It Was Not Enough

Full-year fiscal 2026 revenue was $63.3 billion, up 12%, with non-GAAP EPS of $4.33, up 14%. On any normal reading this is the strongest guide Cisco has issued in a decade, and the $7.5 billion of embedded fiscal 2027 AI infrastructure revenue is a number the company simply did not have a line item for two years ago.

The market’s objection is narrow and it is correct. Cisco’s AI revenue is hardware, sold into hyperscalers who negotiate hard, and it carries a materially lower gross margin than the software and subscription mix that drove the stock’s rerating over the past three years. Growing 15% while gross margin compresses means the company is buying revenue growth with margin, and the market had been paying a software multiple for a business whose fastest-growing segment is now a box business. The stock is not being punished for missing. It is being repriced for what it is becoming.

Our take: the guide is real and the order number is the single best piece of evidence in this note that hyperscaler networking spend is still accelerating into 2027. But we would not chase the dip until there is visibility on where gross margin settles, because the mix shift has at least two more years to run and every incremental dollar of AI revenue makes the blended margin worse. This is a story stock turning back into a cyclical, and the multiple has further to travel than the earnings do.

Coherent: the Print Was Right and the Reaction Was Wrong, and We Owned Half of That

Yesterday this note said Coherent carried more information per dollar of market capitalisation than any other print of the week, because optics volumes lead rack deployments in a way that networking revenue recognised over long contracts does not. That framing held up exactly. Datacenter and communications grew 59% and now accounts for 79% of the company, full-year revenue reached $7.12 billion, up 22.5%, and net earnings swung from a prior-year loss to $805 million. The September quarter guide sits roughly 8% above consensus on revenue and 10% above on earnings. There is nothing in this release to dislike.

The stock is down 5% anyway, and the honest reading is that we got the mechanism right and the reaction wrong for the second consecutive session. The proximate cause is positioning rather than fundamentals. Coherent had already run hard into the print and briefly traded up 8% before erasing it, and the sell-side average target of $415 sits inside a range from $330 to $465, which is a polite way of saying nobody agrees on what this is worth. What we underweighted is that when a stock has already discounted a beat, the beat has to be larger than the one that arrived.

Our take: this remains the cleanest way to own the physical layer of the buildout and the 1.6T transceiver ramp is doing what we said it would. The margin expansion here, against compression at Cisco and Cerebras, is the part of this note we would most want to be right about, because it argues optics is where pricing power actually sits in the interconnect chain.

Cerebras: the Growth Held and the Economics Cracked

Core revenue of $210 million beat the roughly $194 million consensus, core cloud revenue grew 287%, and the company raised full-year core revenue guidance from an $855 million to $865 million base. Those are numbers that would ordinarily lift a second-time reporter.

The guide killed it. Gross margin was pointed at 36% to 38% against 47% in the prior quarter, and the adjusted loss per share came in materially wider than forecast. Nearly ten points of sequential margin compression in a single guide, from a company whose entire investment case rests on wafer-scale silicon being structurally cheaper per token than the Nvidia stack, is close to a thesis event. The stock fell 11.5% after hours and is down roughly 15% now.

Our take: the demand read is real, and an inference cloud that nearly quadrupled matters for anyone modelling accelerator demand outside Nvidia. But a specialised silicon company guiding to sub-40% gross margin while scaling is telling you it is buying share, and buying share against Nvidia is expensive in a way that compounds. Of the three prints, this is the one where the margin line is a signal about competitive position rather than about mix.

July PPI: Flat Headline, Hot Core, Same Energy Trick

Read the release rather than the headline and the composition is nearly identical to yesterday’s CPI. Producer prices were flat only because final demand goods fell 0.7%, driven by a 3.1% drop in energy and a 5.7% decline in gasoline, with food down 0.9%. Strip those out and core goods still rose 0.1%, services rose 0.2%, and the measure the Fed actually watches, final demand less foods, energy and trade services, rose 0.4% on the month against a 0.3% expectation and is running 4.7% year on year. Services excluding trade, transportation and warehousing rose 0.6%, which is the stickiest slice of the entire index.

One detail deserves calling out because it is circular in a way that should make people uncomfortable. The single largest contributor to the services increase was portfolio management fees, up 6.5% on the month. Portfolio management fees are calculated off asset values. In other words, a meaningful chunk of July’s producer services inflation is the equity market rallying, showing up in the inflation statistics, which then feeds a rate decision that moves the equity market. It is a real component and it belongs in the index, but anyone treating it as evidence of underlying demand pressure is reading a mirror.

The market took the headline. The dollar index fell 0.19% to 99.82, the 2-year dropped 4.6bp to 4.16%, the 10-year 4.3bp to 4.66% and the 30-year 2.8bp to 5.23%. Yields fell across the curve, front end leading, which is the textbook response to a softer inflation path.

Our take: for the second day running the market has bought disinflation that was manufactured by an energy component. Gasoline fell 5.7% at the producer level in a month that ended two weeks ago. Since that month closed, Brent has run six consecutive sessions to $88 and the Strait of Hormuz is still not open. September odds on CME FedWatch remain close to a coin flip, with the various calibers we track spread between roughly 40% and 50% and no clean resolution among them. We would fade any move that takes those odds materially below 40% on July data, because the core number in today’s release is the part of it that will still be true in September.

Labour: Claims Tick Up Without Breaking

Initial jobless claims for the week ended August 8 came in at 209,000 against roughly 202,000 expected, an increase of 9,000 from the prior week’s upward-revised 200,000. The four-week moving average was unchanged at 199,000, and continuing claims had been running near 1.80 million.

The right way to hold this is that the level is still historically low and the trend has not broken, but the four-week average sitting exactly at 199,000 while the weekly print pushes above 200,000 for the first time in a month is consistent with the picture the July payrolls report drew, which was hiring that has stopped rather than firing that has started. That distinction matters for the September debate. A committee looking at 209,000 claims does not see an emergency. A committee looking at minus 23,000 payrolls, downward revisions of 103,000 and participation at 61.4% sees something that argues loudly against tightening into it.

Asia: Korea Adds a Third Leg, Tokyo Sets a Record, China Goes the Other Way

The Kospi rose 3.56%, or 234 points, to 6,813.34, a three-week high, led again by the chipmakers. That is the third consecutive session of large gains and follows yesterday’s 3.7% move, which means Korea has added roughly 7.5% in two days on the memory complex. Japan was firm, with the Nikkei up 1.16%, or 784 points, to 68,308 and the Topix setting a record at 4,176.04.

China diverged sharply and it is the more interesting half of the region. The Hang Seng slipped 0.2% to 25,396 despite the regional tech bid, the Shanghai Composite fell 0.5% to 3,926, the Shenzhen Component dropped 0.87%, and materials shares in Hong Kong tumbled 5.4%. Tencent fell 4.5% to a two-week low, and the reason UBS gave is precisely the theme of this entire note: investors are concerned that weak consumption may limit ad-load expansion while AI investments take longer to generate commensurate returns. That is a Chinese platform company being sold for the same reason Cisco and Cerebras are being sold in New York this morning, which is that the spending is visible and the return on it is not yet.

One local detail worth flagging for the optics chain: co-packaged optics names in China opened strong on the Coherent read, with TFC Optical Communication up as much as 13% before momentum faded. The fade is the tell. Even the direct read-through beneficiaries could not hold their gains.

Japanese producer prices rose 7.2% in July, staying high while the Bank of Japan weighs its rate path. A 7.2% producer print in Japan against a 4.7% print in the United States is a reminder that the global inflation impulse has not resolved anywhere near as cleanly as the US front end is currently assuming.

Europe: Records Held, and Crude Did the Work

The Stoxx 600 rose 0.2% to 660.49, holding the record territory it has occupied all week, with European shares supported by the fall in crude shifting the conversation from supply risk to demand. The FTSE 100 fell 0.3%, which is what an energy-heavy index does when Brent drops 1.2%, and the DAX was firm.

The mechanism deserves a note because it cuts both ways. European equity has spent two weeks rallying on semiconductors while its energy and luxury components did the damage. This morning the sign flipped and the index still went up, which means the breadth underneath the record is marginally better than yesterday’s version. We would not overread one session, but a Stoxx 600 that can absorb a crude decline without giving up the level is a slightly different index from the one that needed ASML and Infineon to carry it.

Oil and Geopolitics: the Rhetoric Escalated and the Price Fell

Brent fell toward $87.92, down 1.19%, snapping a six-session advance, with WTI near $81.60. Prices fell despite the rhetoric getting materially worse, and that gap is the thing to explain.

President Trump said the United States has total control over the Strait of Hormuz, told reporters that “we own it, and at some point, maybe they’ll do something, and then they get blown away,” and has demanded reparations from Iran as a condition of talks. Tehran responded by declaring the strait blocked and has continued intermittent attacks on transiting vessels. Negotiations are deadlocked, with both sides hardening, and Iranian officials said the war with the United States could be prolonged.

Against that, two supply-side facts did the work. The IEA’s monthly report put the global market shortfall at 1.8 million barrels a day this quarter, which is bullish, but EIA data showed US crude inventories rose 17.4 million barrels last week, the largest weekly build since early 2023. A build of that size is a demand signal, not a Hormuz signal, and it is what pulled the price down.

Our take: we do not read this as the war premium deflating. The premium is a function of transit counts, and nothing about transit counts improved overnight. A 17.4 million barrel build arriving after six sessions of gains is a positioning event meeting an inventory number, and the structural setup is unchanged. We keep the $85 to $90 band and continue to see the risk skewed higher, because a president claiming ownership of a waterway that the other side says is closed is not a de-escalation, whatever the front-month contract did today.

Tonight: Applied Materials Is the Read on Whether the Order Book Reaches the Toolmakers

Applied Materials reports fiscal Q3 after the close at 4:30pm Eastern, with consensus at roughly $9 billion of revenue and $3.39 of EPS against company guidance of about $8.95 billion plus or minus $500 million. Last quarter it delivered $7.91 billion and $2.86 against $7.68 billion and $2.68 expected.

The setup is unusual. The stock is up 108% year to date and yet sits about 27% below its highs, which is a chart that has already priced both the boom and a scare. The two variables are China and memory, pulling in opposite directions. Applied has guided to $600 million to $710 million of fiscal 2026 revenue lost to tightened US export controls and expects Chinese equipment spending to fall in 2026. Working the other way, the memory capacity expansion behind the entire Korean rally, with Samsung adding roughly 50% of HBM capacity this year and SK Hynix multiplying its infrastructure investment several times over, has to be bought from someone, and Applied is one of a handful of companies it can be bought from.

Our take: this is the cleanest test in the market of whether the order books everybody is reporting have reached the capital equipment layer yet. Given the theme of this morning, watch the gross margin guide at least as closely as the revenue number. If Applied guides margin down while guiding revenue up, that makes four for four, and the margin story stops being about individual mix decisions and becomes the defining feature of this phase of the cycle.

Corporate Headlines Worth the Time

Firefly Aerospace pushed the debut of its Alpha Block 2 rocket to the fourth quarter on its second quarter call, with Flight 8 now expected after acceptance testing, delivery to Vandenberg, static fire and payload integration. Rocket Lab continues to signal that the window for a 2026 Neutron debut is narrowing, having unveiled its GHOST deployable launch system and signed a Neutron launch agreement with Kepler Communications this week. The pattern in launch remains the one we have flagged since July: contracted backlog keeps growing while first-flight dates keep sliding right, and the market has been willing to pay for the first and ignore the second.

SpaceX launched 24 Starlink satellites from Vandenberg on Tuesday. In healthcare, RTW Biotech Opportunities said its portfolio company Avere Therapeutics will merge with NextCure in an all-stock transaction expected to close by year end, a small deal in an M&A tape that has been notably quiet this week.

Key Events Today

• 8:30am: July PPI, released flat month on month against +0.2% expected and 4.7% year on year, with core ex food, energy and trade services at +0.4%
• 8:30am: Initial jobless claims, released at 209,000 against roughly 202,000 expected, four-week average unchanged at 199,000
• Through the day: Fed speakers, the first commentary since both inflation prints landed
• 4:30pm: Applied Materials fiscal Q3 call, consensus roughly $9 billion and $3.39
• Tomorrow: University of Michigan preliminary sentiment and inflation expectations, with retail sales in the coming week
• September 11: August CPI, the last inflation print before the meeting
• September 15-16: FOMC, still close to a coin flip on the calibers we track

What We Are Actually Watching

Three threads.

The first is that the AI trade has changed the question it is asking for the second time in three sessions, and this time the answer is harder. On Tuesday the question was whether the buildout could be funded, and Super Micro and CoreWeave answered it by producing order books too large to argue with. This morning the question is what those orders earn, and Cisco, Cerebras and Coherent answered it from three different layers of the stack with the same word, which is margin. Cisco’s blended gross margin falls as AI hardware grows. Cerebras is guiding ten points of sequential compression to hold share. Coherent is the exception that proves the rule, expanding margin and still getting sold. That is not a bearish development for the cycle. It is the point at which a capex boom stops being a story about who gets the orders and starts being a story about who keeps the profit, and those are very different lists.

The second is that two consecutive inflation releases have been flattered by the same expired energy move, and the market has bought both. CPI on Wednesday was soft because energy fell 1.5% and gasoline fell 2.9% in July. PPI today was flat because energy fell 3.1% and gasoline fell 5.7% in the same July. Meanwhile core PPI ran 0.4% and the stickiest services slice ran 0.6%, and Brent spent the six sessions since that month closed climbing to $88. The front end has rallied on the headline of a report whose core was hotter than expected. We think that trade is on borrowed time, and the August data is what repays it.

The third is that gold at $4,441 has come off yesterday’s $4,493 high on a morning when the dollar fell, yields fell across the entire curve and inflation printed soft. Every one of those inputs should have supported the metal, and it went the other way. Yesterday we flagged that gold was pricing something the front end and a VIX at 15 were not expressing, and the most likely candidate was the shipping lanes. Today gold fell alongside crude, on a day when the rhetoric around Hormuz got worse rather than better. Both moves are consistent with the same explanation, which is that positioning rather than news drove the tape this morning. That is worth knowing before assuming today’s prices contain new information about the war.

Yesterday this note argued that the instrument to watch was no longer the size of the backlog but the cost of carrying it, and pointed at CoreWeave’s interest expense at 25% of revenue. Overnight the market found the same problem one line higher on the income statement. It is not interest that got repriced this morning, it is gross profit, and it got repriced at three companies at once. We will mark that against Applied Materials tonight.

Marks above are pre-market and will move by the open.