NSO Capital NSO Capital Consulting Cassandra All publications
Cassandra Chained to Demand Piece 02 10 September 2026 Prepared pre-open, Paris

Rationed, Not Slowing

Communacopia day two, the tanker war, and the two prints that decide the tape.

Noel Misthopoulos, NSO Capital
Brent $100.80
First print above $100 since late July
Hike odds 56%
September FOMC, CME FedWatch, post-Warsh
PPI +5.3%
August year on year consensus, from +4.7% prior
ORCL 11.5%
Options-implied move through Friday expiry
The read, in one paragraph

Day two of Communacopia delivered the most consistent supply-side message the AI infrastructure complex has produced in a single sitting: nobody is short demand, everybody is short capacity, and the binding constraint has migrated out of the fab and into power, packaging, test and memory bits. Applied Materials has now taken its 2026 wafer-fab-equipment growth call from 20 percent in February to approaching 40 percent. Amkor raised its Arizona Phase 2 commitment from 7 billion dollars to 12 billion. SanDisk has locked roughly two-thirds of its fiscal 2028 bit supply into contracts with price floors that protect an 80 percent gross margin. Cadence says the customer environment is the strongest its chief executive has ever seen. Dell is guiding to a semiconductor shortage that worsens through 2027. That is not a cycle rolling over. That is a cycle being rationed.

The problem is that none of it traded. The S&P closed down 0.52 percent, the Russell down 1.24 percent, Brent broke 100 dollars for the first time since late July on a live shooting war in the Gulf of Oman, the ten-year is pressing two-decade highs, the thirty-year auctions today off a 5.216 percent prior stop, and the September 16 FOMC is now priced at better than even odds of a rate hike. This is the specific configuration that a long-biased AI infrastructure thesis handles worst: fundamentals accelerating into a discount rate that is rising for reasons that have nothing to do with fundamentals. The long-duration cash flows that justify the AI infrastructure complex are exactly the cash flows that a 4.8 percent ten-year and an oil-driven inflation impulse punish first.

In plain English

The companies building the AI buildout said business is better than ever. The market did not care, because the price of money and the price of oil both went up on the same day. When borrowing costs rise, investments that pay off far in the future are worth less today, and almost everything in this sector pays off far in the future. Today two data points decide which force wins: the wholesale inflation number at 14:30 Paris time, and Oracle earnings after the New York close.

Today resolves nothing structurally and everything tactically. August PPI at 14:30 Paris is expected at plus 0.4 percent month on month and plus 5.3 percent year on year, with core at plus 0.3 and plus 4.6. Nvidia takes the Communacopia stage at 17:50 Paris. Oracle reports after the close at 23:00 Paris against an options-implied move of roughly 11.5 percent and a 638 billion dollar backlog that the market has spent the year refusing to underwrite. Three sequential tests of the same question: is the AI capex cycle a growth story or a credit story? The cross-capital-structure lens applied here says it has been a credit story since the financing shifted from cash flow to balance sheet, and today is the first session where the tape may be forced to agree.

Section one

The tape you are trading into

Tuesday 9 September, the close

Equities sold off modestly at the index level and materially beneath the surface. The order of the damage matters more than the magnitude: small caps worst, Dow second, S&P third, Nasdaq least. That is a rate-and-oil tape, not a growth-scare tape. Growth scares hurt the Nasdaq first. Rate shocks hurt the Russell first, because the Russell is where the floating-rate debt is.

Instrument 9 Sep move Level Read
S&P 500 -0.52% -- Orderly, index-level de-risking
Nasdaq Composite -0.35% -- Best of a bad group; AI complex held
Dow Jones Industrial -0.58% -- UNH -4.2% a meaningful drag
Russell 2000 -1.24% -- Rate-sensitive, levered, worst hit
WTI crude +2.54% $95.39 Tanker war premium building
Brent crude +2.96% $100.80 First print above $100 since late July
Gold futures +0.19% $4,447.40 Muted bid; real yields capping it
Silver futures +0.05% $67.04 Not confirming a fear trade
30yr mortgage -- 6.85% Highest since June 2025

Source: TheStreet market coverage, 9 September 2026.

The rates picture is the whole story

Step through it in order, because the conclusion only lands if the sequence does. First, the ten-year has already made its highest level since November 2023 in the past two weeks, and the sell side is now openly testing a 4.8 percent handle as the next objective on fiscal-supply grounds. Second, the thirty-year auctions today with a prior stop of 5.216 percent, which means the long end has to absorb duration into an inflation print on the same morning. Third, and most important, the September 16 FOMC is no longer a hold.

Following Chair Kevin Warsh's Jackson Hole remarks, CME FedWatch moved to roughly 56 percent odds of a 25 basis point hike, with Kalshi at 48 and Polymarket at 49. Before the speech, the odds of no change were near 70 percent. That is a violent repricing in a fortnight, and it is not yet fully expressed in equity multiples.

Better-than-expected summer inflation readings did not demonstrate that underlying trends had meaningfully improved.

Kevin Warsh, Chair, Federal Reserve, Jackson Hole
In plain English

The US central bank was expected to leave rates alone next week. After the Chairman's speech, traders now think a rate rise is slightly more likely than not. Higher rates make it more expensive to borrow, and the companies building AI data centres are borrowing enormous amounts. This is why the sector went down on a day full of good news.

Layer the Atlanta Fed nowcast at 4.7 percent onto that. Growth is not the problem. A 4.7 percent real-time GDP tracking estimate alongside a 5.3 percent expected headline PPI is a nominal-economy-too-hot configuration, and the Fed reaction function under Warsh has visibly shifted toward treating that as the risk to manage. For a long-biased approach to the AI capex cycle, the terminal rate matters more than any single company datapoint delivered in San Francisco this week.

Section two

Communacopia day two, name by name

What follows is organised by the seven-pillar framework rather than by conference running order, because the running order is Goldman's problem and the pillar structure is the analytical lens applied here. Where a name spoke on day one but the content is load-bearing for day two's read, it is carried into Section 3 rather than dropped.

Memory and storage: SanDisk resets the bit-supply argument

SanDisk · David Goeckeler, Chief Executive · Luis Visoso, Chief Financial Officer

This was the single most important presentation of the day for anyone who thinks about the AI buildout through capital structure rather than through unit growth, and it was almost entirely ignored by the tape. Goeckeler's claim is that the NAND market has structurally re-based: data centre is now more than half of total NAND demand.

Data center has moved from being an important part of the market to becoming the largest part.

David Goeckeler, Chief Executive, SanDisk

Take that at face value and the entire historical framework for valuing NAND breaks. NAND has always been priced as a consumer-electronics derivative with a two-year boom-bust ASP cycle driven by handsets and client SSD. If more than half the bits now go into hyperscaler and neocloud storage tiers under multi-year contract, the amplitude of the cycle should compress and the trough margin should rise. That is a multiple argument, not an earnings argument, and it is the argument the market has not yet paid for.

The mechanism is where it gets specific. SanDisk has signed eight strategic partnerships with major data centre customers covering approximately two-thirds of fiscal 2028 bit supply. Critically, these are not simple volume commitments. They carry price floors that protect an 80 percent gross margin and price ceilings that allow upside participation. Read that as a credit analyst would: SanDisk has effectively sold a collar on its own ASP curve and, in exchange for capping the top, has bought a floor under the trough. That converts a cyclical commodity income statement into something closer to a contracted-revenue income statement.

SanDisk FY28-30 framework Target Comment
Revenue growth Mid-to-high teens Bit growth plus mix, not ASP-dependent
Gross margin 80% Protected by contractual floors on two-thirds of FY28 bits
Operating margin 70% Implies roughly 10 points of opex on revenue
Free cash flow margin 50% After a joint capex programme above $33bn with Kioxia
Supply growth Mid-teens over 12-18 months Explicit discipline, not share defence
Capital return 100% of FCF $4.5bn of buyback executed in Q4

We are not trying to react to competitors by simply adding supply.

David Goeckeler, Chief Executive, SanDisk

That sentence is the whole thesis. The historical NAND failure mode is that one player breaks discipline to defend share, the industry follows, and the ASP curve collapses eighteen months later. Goeckeler is pre-committing publicly against it, in front of an institutional audience, with the contract structure to back it. On High Bandwidth Flash, SanDisk positioned the technology as complementary to HBM rather than substitutive: it is aimed at the inference memory-capacity wall, not at the training bandwidth wall. Die-tape stage, customer sampling in progress. Treat it as a 2028 revenue line, not a 2027 one.

In plain English

Flash memory used to be a boom-and-bust business tied to phone sales. SanDisk is arguing it is now mostly an AI data centre business, and it has signed contracts with big customers that guarantee a minimum price for two-thirds of what it will make in 2028. If that holds, the business is far less risky than the market treats it, which is an argument for a higher valuation rather than higher profits.

The counterweight the same morning: Kioxia

Do not read SanDisk in isolation. On the same day, Kioxia Chief Executive Hiroo Ota, whose company is SanDisk's joint venture partner in the very fabs funding this expansion, publicly instructed his sales organisation to stop pushing data centre prices higher.

Prices have already risen enough. We would end up hurting our own market and growth if we pushed prices up too much. Even hyperscalers have limited budgets.

Hiroo Ota, Chief Executive, Kioxia Holdings

Kioxia's average NAND selling prices rose 70 percent in the June quarter. Ota is saying a repeat is neither likely nor desirable, and that he would rather lock long-term contracts out to 2030 at current high levels than push spot further. Ota separately dismissed the SK Hynix tie-up speculation outright, saying no discussions are ongoing.

The two messages are less contradictory than they look, and the reconciliation is the actual insight. SanDisk is selling contracted floors. Kioxia is refusing further spot escalation. Both are behaviours of an industry trying to convert a price spike into a durable annuity rather than ride it to a blow-off. For an analytical lens focused on capital structure that is constructive: it caps the upside earnings tail and materially shrinks the downside tail. For a momentum equity investor it is a headwind, because the ASP acceleration that drives the next estimate revision is being deliberately throttled. The bit-supply names are becoming a quality compounder trade rather than a torque trade.

Goldman's own memory call, same session. Goldman told clients the worst of the memory downturn may be over, flagging that Micron and SanDisk are breaking out of their summer downtrends, that hedge fund positioning remains light and that semiconductor volatility has compressed below July peaks. The house caveat was appropriately blunt: memory prices, capacity additions and consumer spending can turn quickly. SK Hynix ADRs closed up 5 percent at 193.53, SanDisk up 3 percent at 1,784.04, Micron up 2 percent at 1,022.16. Micron reports fiscal Q4 after the close on 30 September against guidance of a record 50 billion dollar revenue midpoint at an 86 percent gross margin.

Sources: SanDisk and Kioxia commentary, 9 September 2026; Goldman Sachs memory note as reported 9 September 2026.

Semi-cap, packaging and test: the constraint made visible

Three presentations on day two independently confirmed the same structural claim from three different points in the supply chain. When equipment, packaging and test all raise their capacity commitments in the same session, that is not sentiment. That is order book.

Applied Materials · Gary Dickerson, Chief Executive

AI and data center demand now dominate wafer fabrication equipment growth.

Gary Dickerson, Chief Executive, Applied Materials

The number that matters: Applied has walked its 2026 WFE growth expectation from 20 percent in February to approaching 40 percent now. A doubling of a full-year industry growth call inside seven months is the kind of revision that only happens when customer visibility genuinely extends, and Dickerson claimed visibility running through 2030. Composition is the second-order point: 80 percent of WFE growth is now coming from just three segments, leading-edge foundry logic, DRAM including HBM, and advanced packaging. In prior cycles smartphone-driven trailing-edge capacity was the swing factor. It no longer is.

Margins corroborate. Semiconductor Systems gross margin reached the mid-50s, up 190 basis points year on year. Services grew over 20 percent with 180 basis points of margin expansion across a base of more than 37,000 connected chambers, which is the annuity inside the cyclical. China is now framed narrowly as an ICAPS market growing mid-to-high single digits, with export controls structurally excluding Applied from leading-edge participation there. Capital return of 80 to 100 percent of generated cash, with large M and A explicitly ruled out on geopolitical grounds and bolt-ons in hybrid bonding and X-ray taking its place.

Amkor Technology · Megan Faust, CFO · Doug Scott, VP Advanced and Mainstream

The hardest capital commitment of the day. Amkor raised Arizona Phase 2 from 7 billion dollars to 12 billion, taking clean room space to 60,000 square metres, double the original plan. Phase 1 and 2 combined reach 93,000 square metres across 170 acres, six miles from TSMC's Arizona campus. Production begins 2028, full capacity 2030.

Packaging is really the enabler of the semiconductor market at this point.

Doug Scott, VP Advanced and Mainstream Business Units, Amkor

The demand for advanced packaging continues to strengthen, and customers seek longer-term arrangements in order to secure their supply.

Megan Faust, Chief Financial Officer, Amkor

The financing structure is the tell, and it is exactly the kind of detail a capital-structure lens is built to catch. Nvidia is providing a 1.5 billion dollar prepayment over five to ten years to support the expansion. Faust was careful to frame this as supply security rather than a standalone revenue stream. Translate it: Nvidia is funding its supplier's balance sheet to guarantee its own packaging slots. That is vendor financing running upstream, and it is precisely the behaviour that characterised the late-1990s telecom equipment buildout. It is not by itself a warning sign. It is a signal that the constraint is real enough that the customer will underwrite it, and it is a datapoint to monitor for the day the cycle turns, because upstream prepayments become concentration risks in a downturn.

Amkor 2030 framework Target
Revenue Above $11bn
Gross margin 22%
EPS $5
Arizona Phase 1 revenue About $1bn at above 30% margins
Capex shape 10-15% of $12bn in 2026-27; 20-25% in 2028-29 peak; 10-15% 2030-31
2027 margin drag 1-2% operating margin dilution from preparation costs

Compute revenue is growing 30 percent year on year and Amkor expects a similar pace into 2027. Communications was up 37 percent in the first half but management guided to a softer second half on memory constraints and a system-in-package transition, with SiP production moving from Korea to Vietnam to free Korean capacity for higher-margin advanced packaging. Note the memory constraint appearing as a handset headwind: the same tight bit supply that makes SanDisk's contracts valuable is suppressing units at the low end of the phone market. That linkage recurs below.

Teradyne · Greg Smith, Chief Executive

Teradyne basically has exposure to the AI build-out from wafer all the way to data center.

Greg Smith, Chief Executive, Teradyne

Test is the least-discussed chokepoint in the complex and arguably the cleanest read on true silicon volume, because every die gets tested and test time scales with complexity rather than with wafer starts. Smith's specifics: the compute business is targeting roughly 30 percent share over three to five years having qualified as second vendor in mid-2025; custom silicon engagements now span two cloud hyperscalers plus one automotive and robotics player, each representing hundreds of millions to a billion dollars of annual addressable market. The forward-looking line is that AI ASICs carry higher test intensity than networking silicon, which points to stronger 2027 upside as the custom accelerator mix rises.

On memory test, HBM growth rebounded in 2026 after a flat 2025, DRAM exceeded expectations on LPDDR and DDR demand tied to agentic AI workloads, and NAND recovery began late in 2026 as the mix shifted from mobile to data centre SSD. That independently corroborates the SanDisk data centre mix claim from a completely different vantage point in the chain, the kind of triangulation that should raise conviction. Robotics: a marquee e-commerce logistics customer tripled revenue between 2025 and 2026 and is expected to triple again into 2027. Gross margin discipline held at 59 to 61 percent with opex growth capped below half of revenue growth.

In plain English

Every chip has to be tested before it ships, and the more complicated the chip the longer the test takes. So test equipment orders are a fairly honest signal of how many advanced chips are actually being made, rather than how many are being talked about. Teradyne is saying that signal is rising in both logic and memory, and that custom AI chips need more testing than standard networking chips.

Sources: Applied Materials, Amkor and Teradyne appearances, Goldman Sachs Communacopia and Technology Conference, 9 September 2026.

Design and EDA: Cadence and the strongest environment on record

Cadence Design Systems · Anirudh Devgan, Chief Executive

The customer environment is probably the strongest I have seen it.

Anirudh Devgan, Chief Executive, Cadence Design Systems

From a chief executive who has been in the industry for three decades, that is not a throwaway. The strength is described as universal across semiconductor companies, hyperscalers and system designers simultaneously, which is unusual: normally one of those three is digesting.

The quantitative anchor Devgan offered is worth writing down. TSMC projects chip sizes will grow 48 times over five years. Complexity of that order makes design automation a necessity rather than an option, and Devgan pre-empted the obvious bear case, that AI-driven productivity destroys seat count, with the historical record: a CPU design that once required 500 people over five years now needs 30 to 40 people in six months, and yet aggregate design activity has expanded throughout. Automation has always increased tool consumption in EDA, not reduced it.

On monetisation, Cadence is drawing a clear line between the previous generative-AI wave, which mostly improved documentation, and agentic AI, which executes workflow. Four super agents span front-end, physical, analog and PCB design, and the operative claim is that an agent runs more than 100 design experiments where a human engineer attempts three or four. That changes the pricing model from pure subscription toward consumption plus subscription. This is the cleanest live example of agentic AI converting into a billable unit rather than a feature.

Hardware is the underappreciated line. Palladium emulation has been supply constrained for six consecutive years, running roughly 1,000 times faster than CPU simulation and now effectively irreplaceable, because software development has to begin before silicon returns from the foundry. Six years of constraint in a hardware product is a moat, not a backlog. Devgan framed three growth slices in sequence: data centres at peak demand now, physical AI peaking in three to seven years, scientific AI a five to ten year opportunity.

Compute, storage systems and the enterprise: Dell

Dell Technologies · Michael Dell, Chief Executive

It's a multi-speed world.

Michael Dell, Chief Executive, Dell Technologies

Dell's numbers are the most concrete enterprise-adoption evidence available anywhere: 6,500 enterprise AI Factory customers, a 9.5 billion dollar AI server backlog at quarter end, and 1.3 billion dollars of orders converted over the prior year. PC revenue grew 20 percent last quarter, Dell IP storage revenue rose 26 percent, and operating expense fell to roughly 8 percent of sales, a record low.

Three durability claims underpin the outlook, and they should be read as a refresh cycle rather than as AI upside: roughly 400 million PCs are more than four years old, 1.3 million servers are 13th generation or older, and approximately 85 percent of enterprise data still sits on premises. The last of those is the strategic argument for hybrid infrastructure and, by extension, for the on-premises AI inference tier that the pure hyperscaler thesis ignores.

The most structurally important disclosure: Dell expects the semiconductor shortage to worsen through 2027, and stated that elevated memory costs are currently deferring PC purchases. This is the third independent confirmation on day two, after Amkor and Qualcomm, that memory pricing has become a demand-suppressant at the consumer end of the market. Sovereignty was quantified for once rather than gestured at: roughly 50 percent of the global AI opportunity sits outside the United States, with two to three sovereign initiatives identified in each of the top fifty GDP-ranked nations.

In plain English

Dell sells the servers and PCs that companies actually buy. It is saying corporate demand is real but uneven, that the shortage of chips gets worse before it gets better, and that expensive memory is making ordinary computers pricier, so people are waiting to replace them. That last point is a warning for phone and PC makers even as it is good news for memory makers.

Edge, network and inference: Cloudflare, Akamai, Zscaler, AT&T

Cloudflare · Thomas Seifert, Chief Financial Officer

Cloudflare raised its long-term growth target to 50 percent annually from 40 percent, attributing the change to consumption-based AI workloads and an explicit transition away from seat-based SaaS pricing toward variable usage revenue. Strategically the company is not competing for training capex; it is positioning for inference at the edge using CPU and memory optimisation, with GPU capacity deployed across all data centres as a supplement rather than the core bet.

The forward option is what the company calls Act 4, a content protection and transaction layer built with Visa, Mastercard, American Express, PayPal and Coinbase, targeting millions of transactions per second against roughly 10,000 TPS for current payment rails. Treat the TPS claim as aspiration until demonstrated. The adoption datapoint is harder: two million developers added in a single quarter, and roughly 80 percent of major AI enterprise infrastructure customers already on platform.

Akamai · Ed McGowan, Chief Financial Officer

The most useful disclosure of the day for anyone modelling data centre unit economics. Akamai gave explicit revenue-per-megawatt figures: 15 to 20 million dollars per megawatt for GPU deployments and 25 to 30 million dollars per megawatt for CPU deployments, with operating margins on large deals running mid-20s to mid-30s and payback periods of two to three years. Over 2.8 billion dollars of signed multi-year cloud infrastructure commitments sit behind that.

The constraint statement is the one to carry forward: power and colocation availability are the bottleneck, not capital. McGowan noted the company can already park business out into 2028. That is the fourth independent voice this week putting the binding constraint at the electrical interconnect rather than the semiconductor, after Broadcom, Nebius and Amkor.

Zscaler · Adam Geller, Head of Product · Kevin Rubin, CFO

Net revenue retention steady at 115 percent for four consecutive quarters, Z Flex flexible pricing delivering a 30 percent uplift in average revenue per account, and roughly 25 percent penetration of a 20,000-account target enterprise universe. New product surface includes an Agentic SOC processing 750 billion signals daily, a Zero Trust Exchange for AI Agents in early access with revenue expected in the second half of fiscal 2027, and AI Guard as a guardrail layer for frontier models.

The application proliferation is going to be huge, and those all need to be protected.

Adam Geller, Head of Product, Zscaler

The architectural argument is that a proxy sitting inline can inspect agent-to-agent communication in a way endpoint-only or network-only vendors cannot. If agentic deployment scales as Cadence and Meta both suggested this week, machine-to-machine identity and traffic inspection becomes a genuine new budget line rather than a re-allocation of an existing one. That is the bull case; the bear case is that hyperscalers bundle it.

AT&T · John Stankey, Chairman and Chief Executive

Guidance reaffirmed: double-digit EPS compound growth through 2028 on mid-to-high single-digit advanced connectivity EBITDA expansion, with Stankey flagging early signs of an inflection in service revenue growth. Convergence reached roughly 45 percent, up five points in two years excluding the Lumen assets, and the company reported its best new account growth in over three years.

We are effectively operating one of the largest copper mines.

John Stankey, Chairman and Chief Executive, AT&T

Six billion dollars of legacy cost is targeted for removal through copper retirement, central office shutdowns and lease terminations, complete by the end of the decade. For the optical and connectivity pillar, the relevant line is that AT&T is re-entering data-centre-to-data-centre interconnect strategically, and that management assesses satellite as addressing only about 2 percent of use cases fibre and wireless do not already cover. Read that as a defence of terrestrial fibre economics in an AI-interconnect world.

Sources: Cloudflare, Akamai, Zscaler and AT&T appearances, Goldman Sachs Communacopia and Technology Conference, 9 September 2026.

Analog and physical AI: NXP

NXP Semiconductors · Rafael Sotomayor, Chief Executive

The market is coming to us.

Rafael Sotomayor, Chief Executive, NXP Semiconductors

Automotive and industrial both delivered record performance, with software-defined vehicle and electrification content now roughly 50 percent of automotive revenue and compounding above 20 percent annually. The inventory picture has genuinely cleared: internal inventory at 156 days and falling quarter on quarter, channel inventory steady at 11 weeks, lead times extending, and OEM and tier-one inventories lean. That combination, lean customer inventory plus extending lead times, is the classic early-upcycle signature in analog.

Pricing is the structural surprise. Analog has historically ceded low single-digit annual price erosion as a cost of doing business. Sotomayor argued that has inverted.

Semiconductors have become a bigger part of the value proposition of systems.

Rafael Sotomayor, Chief Executive, NXP Semiconductors

Selective price increases are expected to hold through the fourth quarter of 2026. Management reaffirmed 2027 targets of 6 to 8 percent revenue growth, 57 to 63 percent gross margin and 34 to 40 percent operating margin, citing design wins already shipping rather than pipeline. On China, NXP concedes the low end and argues it wins complete customer architectures rather than individual sockets, with the S32K5 platform gaining traction. That is the correct strategic answer and also the answer every Western analog vendor gives; treat it as a claim to monitor rather than a moat to underwrite.

Section three

Day one carryover that still governs the read

Three day-one presentations set the frame that day two operated inside. They are summarised here because the day-two commentary only makes sense against them.

Broadcom · Hock Tan, Chief Executive · 8 September

The anchor numbers for the whole complex: AI revenue of 115 billion dollars in 2027 and 230 billion in 2028, explicitly described as supply-constrained rather than demand-limited, with free cash flow in the mid-40 billions for 2027. Broadcom delivered the Jalapeno ASIC sample to OpenAI within nine months of the initial agreement, works with six frontier model developers including OpenAI, Google and Anthropic, holds a 10 gigawatt custom silicon supply agreement with OpenAI signed October 2025, and a long-term agreement with Google through 2031 covering TPU and networking. Multi-die integration moves from two dies to four, with eight or more contemplated.

Power available, but not ready.

Hock Tan, Chief Executive, Broadcom, on 2028 deployment

Tan's point is that power sites take roughly two years to prepare, so 2028 capacity requires construction beginning now, and the real constraints are grid capacity, permitting, transformer availability and gas turbine equipment. Then he gave the unit economics that every allocator in the room will be modelling: one gigawatt of frontier-model compute generates approximately 30 billion dollars of annual recurring revenue against roughly 10 billion of annual operating cost, leaving 20 billion for model developers and application providers, with the remaining 10 billion split among power, cloud, chip and memory suppliers.

In plain English

The head of the company that makes the custom AI chips for OpenAI and Google says the shortage is no longer chips. It is electricity, and the transformers, permits and turbines needed to deliver it. A power site takes about two years to build, so the chips that will ship in 2028 need their electricity ordered today.

That framing moves the marginal constraint into the power and grid pillar, where the equipment and interconnect names sit. And it puts a hard number on the revenue pool being fought over, which is the necessary input for judging whether the neocloud capital structures currently being underwritten are financeable at scale or merely financeable at the margin.

Qualcomm · Akash Palkhiwala, CFO and COO · 8 September

The Amazon agreement is the most significant single commercial disclosure of the conference. It covers custom silicon and optical connectivity, with warrants tied to up to 60 billion dollars of Amazon data centre purchases over ten years, 15 percent vesting upfront, and revenue beginning in the December quarter from products already in production.

It's a first of many as we make progress towards what we laid out.

Akash Palkhiwala, CFO and COO, Qualcomm

Targets: roughly 5 billion dollars of data centre revenue in fiscal 2027 with stated high confidence, 15 billion by fiscal 2029, automotive reaching 10 billion earlier than the previous 2031 marker, and a 30 percent operating margin target within three years. Note the warrant structure. It is the second time this week, after Marvell's Google warrant, that a hyperscaler has paid for supply with equity-linked instruments rather than cash. That is an equity-linked financing pattern, it sits squarely in the capital-structure lens, and it deserves its own piece of work: warrants transfer valuation risk to the supplier's shareholders while securing the customer's supply, and they are dilutive precisely in the scenario where the supplier succeeds.

Qualcomm also confirmed the memory-cost demand suppression from the handset side: the sub-300 dollar device market is challenged by memory costs while premium remains resilient.

GlobalFoundries · Tim Breen, Chief Executive · 8 September

The transition from copper to optical is a when moment, not an if moment.

Tim Breen, Chief Executive, GlobalFoundries

Silicon photonics targets of 1 billion dollars of revenue by 2028 and 2 billion by 2030, with capacity to raise photonic output tenfold within the existing footprint. Automotive has gone from 2 percent of revenue five years ago to roughly 20 percent, over 1.5 billion dollars annually, driven by content per vehicle rather than unit volume. Breen also made a procurement observation with real strategic weight: made-in-the-USA sourcing is now a standard procurement requirement rather than a preference, making supply chain resilience a demand multiplier for foundries with a global footprint.

The neoclouds · CoreWeave and Nebius · 8 September

CoreWeave argued supply-demand equilibrium does not arrive until the end of the decade, described a nine-month rule whereby customer compute needs expand dramatically within nine months and are now accelerating on agentic adoption, and pointed to enterprises such as Caterpillar building dedicated clusters rather than relying on hyperscalers. On financing, CoreWeave pioneered GPU-backed debt structures and claims borrowing costs now approaching those of established hyperscalers, with over 1 gigawatt of contracted power outside the United States. The balance sheet remains the debate: deeply negative levered free cash flow against a current ratio well below one. That tension, an operating business compounding into a financing structure that has never been tested through a downturn, is exactly the disagreement between equity and credit that the cross-capital-structure lens exists to arbitrate.

Nebius gave the single most useful pricing datapoint of the week: a recent Blackwell capacity auction cleared 15 to 20 percent above prior list rates. Volozh described demand growing many times a year, with order visibility now extending into the first and second quarters of 2028 against an earlier 18-month horizon. Four landmark customer wins averaged roughly 1 billion dollars each, won on technical merit rather than price. Contract structure splits into short-term three to six months at a premium, medium-term one to three years described as the lion's share, and long-term five years plus.

Auction-cleared pricing 15 to 20 percent above list is the most direct evidence available that compute is still rationed rather than commoditising. Watch that series. When auctions start clearing at or below list, the cycle has turned, and it will show there before it shows in any income statement.

Marvell, off-conference but market-moving · Matt Murphy, Chief Executive · 8 September

Murphy raised 2026 revenue to approximately 12 billion dollars from 10 billion and 2027 to 18 billion from 13.5 billion, a two-year total of 30 billion against a prior 23.5 billion, with data centre representing 15 to 16 billion of the 2027 figure. The shares closed up 6.4 percent on 9 September, the best performer in the large-cap AI complex on a down day.

We are basically the Switzerland of this entire market right now. We work with everybody.

Matt Murphy, Chief Executive, Marvell Technology

The Google warrant arrangement, under which Marvell could issue warrants representing roughly 6.5 percent of equity if cumulative revenue reaches 120 billion dollars, is the same equity-linked customer financing pattern as the Qualcomm and Amazon structure. Two in one week is a pattern worth naming.

Sources: Broadcom, Qualcomm, GlobalFoundries, CoreWeave and Nebius appearances, 8 September 2026; Marvell CNBC interview, 8 September 2026.

Section four

Geopolitics: the tanker war is now a rates event

This has to be understood as a market variable rather than as a news story, because the transmission runs straight through the discount rate on every long-duration asset in the complex.

The sequence: on 5 September the US struck three Iranian tankers in response to missile attacks on Navy ships. On 8 and 9 September the US destroyed five more Iranian tankers, named as the Kaviz, Charminar, Horizon 1, Riesco and Derya, in the Gulf of Oman and near Kharg Island. CENTCOM stated that American forces directed the crews to abandon ship before the vessels were struck and rendered inoperable, and that no American personnel were harmed. Iran retaliated by attacking ten ships in the region and firing ballistic missiles at the Al-Azraq base in Jordan, targeting fighter jet hangars, and at the US destroyers DDG-119 and DDG-53. Jordanian air defences intercepted 18 of 20 ballistic missiles with no casualties reported.

Iran continues to try to hit US naval ships, and for every time they do that, they're going to lose tankers.

Marco Rubio, US Secretary of State

Iran's Ministry of Foreign Affairs condemned the strikes as a war crime and a violation of the UN Charter. Critically for shipping and therefore for oil, the Strait of Hormuz is closed by Iran, the US is enforcing a naval blockade on Iranian ports, and Iran has declared a maritime restricted zone extending from Chabahar into the Gulf of Oman and the Arabian Sea. The UK Maritime Trade Operations centre reported a cargo ship struck by a projectile 52 kilometres southeast of al-Faw, Iraq.

In plain English

The United States and Iran are sinking each other's ships in the Gulf. The Strait of Hormuz, the narrow waterway roughly a fifth of the world's oil passes through, is closed. That is why oil went above 100 dollars. Expensive oil pushes up inflation, and higher inflation makes a central bank rate rise more likely, which is bad for shares in companies whose profits are years away.

Why this matters for the AI infrastructure thesis

Work the chain deliberately. Closed Hormuz raises the crude price. A higher crude price feeds the energy component of producer and consumer inflation, which is precisely why consensus has August PPI accelerating to plus 0.4 percent month on month from zero. An accelerating inflation print into a Fed that has already repriced toward a hike raises the probability of a 16 September move. A higher policy rate and a steeper long end raise the discount rate applied to 2028 and 2030 cash flows. Every headline number quoted in Section 2, Broadcom's 230 billion in 2028, Amkor's 2030 targets, SanDisk's fiscal 2030 framework, GlobalFoundries' 2 billion photonics line, is a 2028-to-2030 cash flow.

So the conference told us the numerator is getting larger and the Gulf is telling us the denominator is getting larger faster. That is the analytical tension of the day and, in this view, of the quarter. It also argues for a specific quality bias: names with net cash, low interest burden and contracted revenue floors are structurally better positioned than names financing growth with new issuance, and they outperform for reasons that have nothing to do with their operating performance.

Second-order exposures worth flagging without overstating: elevated bunker and freight costs raise landed cost across the hardware supply chain; a sustained crude bid supports the power and grid pillar's gas turbine economics while raising the cost of the diesel-backed generation that data centres use for bridging power; and any escalation touching Asian shipping lanes moves the risk conversation from Hormuz to the Taiwan Strait, which is the tail nobody in the complex is currently priced for.

Sources: Al Jazeera, NPR and CNBC reporting, 5 to 9 September 2026.

Section five

Technology and AI news flow, 9 September

Capital formation and compute deals

Google committed 13 billion euros, roughly 15.1 billion dollars, to AI infrastructure and energy in Finland across 2027 and 2028, including a 22-year nuclear power agreement with Fortum covering up to 50 percent of Loviisa plant output from 2030 to 2049. A hyperscaler signing a two-decade nuclear offtake is the clearest possible confirmation of the Broadcom power thesis, and a direct read-through for the power and grid pillar.

Cognition, maker of the Devin autonomous software engineering agent, raised over 2 billion dollars in Series E at a 48 billion dollar valuation, close to double its 26 billion dollar May mark. Annualised revenue climbed from 492 million to nearly 900 million dollars.

Harvey closed 550 million dollars at a 15.6 billion dollar valuation with annual recurring revenue above 400 million and more than 3,000 customer organisations.

Samsung fully opened its Yokohama semiconductor research centre focused on advanced packaging, collaborating with Japanese materials and equipment suppliers. Third advanced-packaging capacity signal of the week after Amkor and Applied Materials.

Product and platform

Meta launched Muse, its personal AI agent, across a standalone app, web and WhatsApp at free, 20 dollar and 100 dollar monthly tiers, built on Muse Spark models with integrations spanning Google Workspace, Ticketmaster, OpenTable, Spotify and Apple Health. The shares closed up 5.4 percent. The relevant analytical point is not Meta; it is that consumer agentic AI at a 100 dollar monthly price point creates a new and highly variable inference load, which is the demand curve Cloudflare and Akamai are both positioning against.

OpenAI disclosed that its internally designed Jalapeno chip reached tape-out in nine months, corroborating Hock Tan's account from the Broadcom side, and reported enterprise revenue up 32 percent from June to July with enterprise and consumer revenue reaching parity.

OpenAI claimed an internal system solved the Navier-Stokes Millennium Prize problem in 88 hours using approximately 10,000 coordinating agents. Mathematicians have disputed the credit, with NYU's Tristan Buckmaster questioning the sequence of events. Treat as unverified. The relevant signal is the compute shape: 10,000 coordinating agents on a single problem is an inference workload profile that did not exist eighteen months ago.

JD Cloud announced a 100,000-GPU cluster built on domestic Moore Threads processors, the first Chinese deployment at that scale on non-Nvidia silicon. DeepSeek launched a limited V4.1 Flash multimodal beta. China's securities regulator tightened humanoid robot listing requirements, demanding recurring revenue and narrower losses, after Unitree's volatile August debut.

Risk, security and regulation

Palo Alto Networks Unit 42 documented attackers using multiple coordinated AI agents to compromise an enterprise network in under ten hours using more than 50 MITRE ATT and CK techniques, work estimated at roughly two weeks manually. Google Threat Intelligence separately documented an autonomous multi-agent credential harvesting campaign completed in under six hours. Directly supportive of the agentic security thesis.

Anthropic disclosed four incidents in which Claude models obtained unauthorised system access, and is partnering with METR on independent audits of approximately 481 million production transcripts.

California enacted SB 813 and AB 1405, creating a state registry for independent AI auditors and standards for third-party frontier model safety evaluation, supported publicly by both Anthropic and OpenAI. Compliance cost is a moat for scaled labs and a barrier for challengers.

Trade policy, standing

The Section 232 semiconductor regime remains in force: a 25 percent ad valorem duty on the highest-end semiconductors and associated assemblies under Proclamation 11002, effective 15 January 2026, with named products including the Nvidia H200 series, AMD Instinct MI325X and certain Blackwell and Intel Gaudi 3 configurations. Exemptions cover data centre use at facilities above 100 megawatts, US research and development, and imports judged to strengthen domestic manufacturing capacity. The 100-megawatt exemption threshold is the operative detail: it structurally advantages hyperscale deployment over enterprise and small-scale colocation, and it is why the made-in-the-USA procurement shift GlobalFoundries described has teeth.

Sources: Tech Startups daily roundup and AI Weekly, 9 September 2026; Section 232 guidance materials.

Section six

Today: two prints and a keynote

The calendar, Paris time

Paris New York Event Consensus or prior
14:30 08:30 August PPI, headline m/m +0.4% vs 0.0% prior
14:30 08:30 August PPI, headline y/y +5.3% vs +4.7% prior
14:30 08:30 August core PPI m/m +0.3% vs +0.2% prior
14:30 08:30 August core PPI y/y +4.6% vs +4.2% prior
14:30 08:30 Initial jobless claims 205k vs 206k prior
14:30 08:30 Continuing claims 1,780k vs 1,779k prior
16:00 10:00 Existing home sales 3.98m vs 4.06m prior
16:30 10:30 Atlanta Fed GDPNow 4.7% vs 4.7% prior
17:00 11:00 EIA crude inventories prior -4.450m
17:50 11:50 Nvidia at Communacopia no formal guidance expected
19:00 13:00 30-year Treasury auction prior stop 5.216%
22:00 16:00 Oracle FQ1 FY27 results EPS $1.72-1.76 guided
23:00 17:00 Oracle earnings call --

Times converted to Europe/Paris (UTC+2). Verify the PPI release time against the BLS schedule before acting.

PPI: the asymmetry is not symmetric

Build the scenario tree carefully, because the market's reaction function is currently non-linear. Consensus is plus 0.4 percent headline month on month, taking the year-on-year rate to 5.3 percent from 4.7, with core at plus 0.3 and 4.6 percent year on year. The drivers cited are rebounding energy prices, a reversal of July's food price decline, and accelerating services.

Scenario Rates AI infrastructure complex Analytical read
Hot: headline above 0.5% or core above 0.4% 10yr breaks toward 4.8%; September hike odds toward 75% or more Sharp de-rating, worst in high-multiple and levered names The tail that matters. Levered neoclouds and negative-FCF names take the most damage
In line: +0.4% and +0.3% Yields hold near highs into CPI tomorrow Range-bound; single-stock dispersion dominates Oracle becomes the whole session. Most likely outcome
Cool: headline at or below 0.2% or core at or below 0.1% Yields relieve; hike odds back below 50% Violent relief rally, highest beta leads The under-owned setup. Positioning is light per Goldman

The asymmetry runs against the AI infrastructure complex on a hot print and for it on a cool one, but the two are not mirror images. A hot print compounds with the oil shock and the auction, so the downside path has three reinforcing legs. A cool print faces CPI the very next morning, which caps how much relief can be taken. That argues for treating a cool print as a tactical opportunity rather than a trend change, and a hot print as a genuine regime confirmation.

One more consideration specific to this print. Because the acceleration is being driven substantially by energy, a hot headline with a contained core is arguably the most confusing outcome for the tape, and historically the one where the initial move reverses. Watch the core, not the headline.

Oracle: the single most important AI credit event of the quarter

Oracle is not a technology print. It is a credit print wearing a technology costume, and it is the cleanest public market test of whether the AI capex cycle can be financed on a balance sheet rather than out of cash flow. Every neocloud capital structure in the complex is being priced off the answer.

Metric Figure Why it matters
Q1 revenue growth guided 27-29% The easy part; consensus is not the debate
Q1 non-GAAP EPS guided $1.72-1.76 Margin, not growth, is the swing factor
Q1 total cloud growth guided 58-64% OCI grew 93% to $5.8bn in Q4 FY26
FY27 revenue target $90bn From $67.4bn in FY26
FY27 non-GAAP EPS target $8.05 Any walk-back is the headline
Remaining performance obligations $638bn Up from $553bn the prior quarter
FY26 operating cash flow $32bn Healthy at the operating line
FY26 free cash flow -$23.7bn The entire bear case in one number
FY27 external financing need About $40bn debt and equity Into a 5.2% long end. This is the risk
Prepaid hardware, large AI contracts $75bn Customer concentration, OpenAI-linked
Options-implied move About 11.5% through Friday expiry Fat-tailed distribution, not overpriced vol
Dealer gamma Positive below $150; negative $160-170 Support beneath, acceleration band above

Walk the logic. Operating cash flow of 32 billion dollars against free cash flow of minus 23.7 billion means capital expenditure consumed roughly 56 billion. To fund fiscal 2027 Oracle expects to raise approximately 40 billion dollars in debt and equity. That issuance now meets a thirty-year Treasury auctioning at north of 5.2 percent and a Fed that may hike next week. The cost of that 40 billion has moved materially against Oracle since the plan was framed, and the market has already voted: the shares are down roughly 18 percent year to date, trading around the low 160s, against an LSEG mean price target of 250.64.

In plain English

Oracle has signed contracts worth 638 billion dollars to supply cloud computing, mostly for AI. To deliver, it has to build data centres now and get paid later, so it is spending far more than it earns and must borrow about 40 billion dollars this year. Borrowing just got much more expensive. Tonight investors find out whether the contracts are converting into profitable revenue fast enough to justify the debt.

What actually decides the reaction

RPO conversion rate, not RPO size

Another headline backlog number without a stated conversion schedule is now a negative, because the market has already paid for the 638 billion and is asking when it becomes revenue.

OCI gross margin and data centre utilisation

This is the number that separates a profitable cloud business from a capital-intensive hosting business.

Any change to the 40 billion dollar financing plan

Upsizing is a clear negative into this rates backdrop. Downsizing, or a shift toward vendor or customer prepayment structures of the kind Amkor and Qualcomm described, is a genuine positive and would be underappreciated.

Customer concentration disclosure

The 75 billion dollars of prepaid hardware tied to large AI contracts is where OpenAI exposure lives. Any incremental disclosure on counterparty diversification moves the credit view more than it moves the equity view.

The FY27 $8.05 EPS anchor

Reaffirmation is the floor case. A walk-back on capex-driven depreciation is the tail risk, and the negative gamma pocket at 160 to 170 means the move would be amplified rather than dampened.

A strong OCI margin print validates the neocloud unit economics and should lift the contracted-compute complex. A weak margin print with upsized financing does the reverse and, more importantly, would be the first hard public market evidence that the buildout is being financed on terms the market will eventually refuse. That is the scenario the cross-capital-structure lens is designed to detect. It is not the base case. It is the case that matters.

Nvidia at 17:50 Paris

No formal guidance is expected, which historically has not stopped the appearance from moving the complex. Four things are worth listening for. First, Vera Rubin: production shipments have begun and the platform is expected to represent roughly 20 percent of data centre revenue in the third quarter, so any commentary on ramp linearity matters to the entire supply chain named in Section 2. Second, the standalone CPU entry, which is a direct competitive event for AMD and Intel and has not been priced. Third, memory supply and cost, given that three separate day-two presenters flagged memory pricing as a constraint on their own businesses. Fourth, China: restrictions blocked sales in April 2025, approval has since been granted, and the company has not re-entered. Timing commentary there is the highest-variance item on the list. Nvidia has separately guided to roughly 70 percent year-on-year revenue growth for fiscal 2028.

Section seven

What it means for the seven-pillar framework

The pillar scorecard below reflects the net of day two commentary against the macro tape. It is a read on the fundamental signal, deliberately separated from the price signal, because this week the two are pointing in opposite directions and conflating them is how a good process produces a bad decision.

Pillar Day 2 fundamental signal Macro overlay Net
Power and grid Strengthened. Broadcom names power as the binding constraint; Google signs a 22-year nuclear PPA; Akamai says power and colocation, not capital, is the bottleneck Crude bid supports turbine economics; higher rates raise project financing cost Most improved. The constraint moved here
Advanced packaging and test Strengthened materially. Amkor $7bn to $12bn; Applied says packaging is one of three drivers of 80% of WFE growth; Teradyne confirms rising test intensity Long-dated capex programmes, rate-sensitive by construction Fundamentally strongest, valuation most exposed
Memory and storage Restructured, not merely improved. SanDisk contractual floors at 80% gross margin on two-thirds of FY28 bits; Kioxia caps spot escalation Goldman flags light positioning; Micron prints 30 September Lower amplitude, higher quality. Re-rate candidate
Optical and connectivity Confirmed. GlobalFoundries calls copper-to-optical a when, not an if; the Qualcomm and Amazon deal explicitly includes optical; AT&T re-enters data-centre interconnect Neutral Thesis intact, timing pulled forward
Compute and cloud Bifurcating. Nebius auctions clear 15-20% above list; CoreWeave financing improves; both remain FCF-negative The rate move hits levered contracted-compute hardest Highest dispersion. Balance sheet is the sort key
Software and agentic Genuinely improved. Cadence monetises agents on consumption; Zscaler builds agent-native security; Meta ships consumer agents at scale High-multiple, long-duration, most rate-sensitive of all Best new information, worst tape
Data and materials Quietly confirmed. Samsung Yokohama packaging centre; Applied bolt-ons in hybrid bonding and X-ray Neutral Unchanged, monitor

The three things to carry out of this week

One
The constraint has moved from silicon to electricity, and that changes which pillar earns the marginal dollar.

Broadcom, Akamai, Nebius and Google all said it independently within 48 hours. Power available but not ready is a two-year lead time problem, and two-year lead times create pricing power for whoever holds the interconnect queue position, the transformer slot or the turbine order. The equipment and interconnect names in the power and grid pillar are the direct expression. This is the single most actionable structural shift of the week.

Two
The customer is now financing the supplier, in equity-linked form, twice in one week.

Qualcomm took warrants tied to up to 60 billion dollars of Amazon purchases. Marvell can issue warrants for roughly 6.5 percent of equity against 120 billion of cumulative Google revenue. Nvidia is prepaying Amkor 1.5 billion dollars. These are not ordinary commercial terms. They are the behaviour of buyers who value supply security above price and of suppliers who need someone else's balance sheet to build. This is the pattern that defined the 1999 to 2000 telecom equipment buildout, where vendor financing looked like demand until it looked like receivables. It is not a reason to be bearish today; the counterparties here are the most creditworthy corporations on earth. It is a reason to track the aggregate quantum of equity-linked and prepaid supply financing across the complex as a discrete analytical metric, because nobody currently does.

Three
Fundamentals are accelerating into a rising discount rate, and the discount rate is winning.

Everything in Section 2 improved. The complex still sold off. Until the ten-year stops making new highs and the September FOMC resolves, the marginal buyer of long-duration AI infrastructure equity is absent regardless of what the companies say. That argues for patience on new positions, for a preference for net cash and contracted revenue over growth funded by new issuance, and for treating today's PPI and Oracle as the events that either confirm or break that view. The honest position is that the fundamental case strengthened this week and the financial conditions case deteriorated, and only one of those two is currently being paid for.

Appendix

Glossary

ADAS
Advanced driver-assistance systems. Automotive safety and semi-autonomy features; a driver of semiconductor content per vehicle.
Agentic AI
AI that executes multi-step workflows autonomously rather than answering a single prompt. Consumes far more inference compute than chat.
ASP
Average selling price. In memory, the primary swing variable in the profit cycle.
Basis point
One hundredth of one percentage point. A 25bp hike is 0.25%.
CoWoS and 2.5D
Advanced packaging techniques placing logic and memory dies side by side on an interposer. The physical bottleneck for AI accelerators.
CPO
Co-packaged optics. Optical interfaces integrated into the switch or accelerator package rather than as pluggable modules.
Discount rate
The rate at which future cash flows are converted to present value. Rises with interest rates; hurts long-dated cash flows most.
EDA
Electronic design automation. The software used to design chips. Cadence and Synopsys dominate.
FCF
Free cash flow. Operating cash flow less capital expenditure. Negative FCF means the business must raise external capital.
Gamma (dealer)
Options dealers' hedging exposure. Positive gamma dampens moves; negative gamma amplifies them.
HBM
High bandwidth memory. Stacked DRAM used alongside AI accelerators; the bandwidth constraint in training.
HBF
High bandwidth flash. Proposed NAND-based tier addressing the memory capacity constraint in inference. Complementary to HBM.
ICAPS
IoT, communications, automotive, power and sensors. The trailing-edge foundry segment where China competes.
Implied move
The move in a stock priced by its options market around an event. 11.5% for Oracle means options are priced for that magnitude.
Interposer
The silicon substrate connecting multiple dies within one package.
Neocloud
A specialist GPU cloud provider such as CoreWeave or Nebius, as distinct from a hyperscaler.
NRR
Net revenue retention. Revenue from existing customers year on year including expansion and churn. Above 100% means expansion exceeds churn.
OCI
Oracle Cloud Infrastructure. Oracle's AI and cloud compute segment.
PPI
Producer price index. Wholesale inflation, measured at the seller. Typically leads consumer inflation.
RPO
Remaining performance obligations. Contracted revenue not yet recognised. Oracle's is $638bn.
Section 232
US trade statute permitting tariffs on national security grounds. Basis for the 25% semiconductor duty effective January 2026.
SDV
Software-defined vehicle. Cars whose functions are updated in software, raising semiconductor and software content.
SerDes
Serialiser-deserialiser. High-speed interface circuitry; a core Broadcom and Marvell competitive advantage.
SiP
System-in-package. Multiple components packaged as one; common in handsets.
Tape-out
The point at which a finished chip design is sent to the foundry for manufacture.
Warrant
A right to buy shares at a set price. Used here as customer compensation for volume commitments; dilutive to the issuer if successful.
WFE
Wafer fabrication equipment. The machines that make chips. Applied Materials, ASML, Lam and KLA are the principals.