The incremental signal from Friday’s trading is arguably more useful than Thursday night’s earnings headlines. The market is no longer rewarding “AI exposure” indiscriminately. Marvell Technology, Rubrik and SentinelOne delivered fundamentally respectable results and were sold; Elastic and Workday were rewarded. That tells us investors are beginning to differentiate between visible incremental AI revenue, distant optionality and already-capitalised expectations. At the same time, the infrastructure story continues to lengthen through SK Hynix and Anthropic/Nscale, while OpenAI’s decision to cut Cursor off from its models introduces a new strategic risk for application companies built on third-party frontier models.
1. Marvell Technology: Friday’s >8% sell-off is the most important semiconductor datapoint after Nvidia because the market is explicitly distinguishing between AI demand and when shareholders actually get paid.
Marvell’s underlying print was strong — Q2 revenue of $2.739bn, +37% yoy, with data-centre revenue +46%, alongside a higher multiyear outlook — but investors focused on management indicating that substantial revenue from the new Google custom-chip relationship does not begin until roughly fiscal 2029. Reuters reported shares falling >8%, potentially erasing c.$17.4bn of market value, despite at least eight brokerages raising price targets. The debate has therefore changed materially. The bull case remains compelling: Marvell is increasingly positioned across custom accelerators, connectivity, optical DSP and SerDes, and management expects roughly 45% FY27 revenue growth and c.$18bn of FY28 revenue. But after the stock nearly tripled this year, investors are no longer willing to capitalise a very large 2029–33 Google opportunity immediately. This is an important read-through for Broadcom, Credo, Astera Labs, Arista Networks, Coherent and Lumentum: custom silicon remains structurally attractive, but expectations around hyperscaler design wins have become sufficiently elevated that timing is now almost as important as TAM. It is also a warning for Nvidia bulls — extraordinary end-market demand does not guarantee every AI semiconductor equity continues rerating.
2. Elastic’s c.20%+ post-results rally is perhaps the cleanest proof this week that investors will pay for software where AI is visibly expanding the underlying workload rather than threatening the pricing unit.
Elastic reported revenue of $478m, +15% yoy, but the more important figures were sales-led subscription revenue +18%, cRPO +21% to $1.153bn, and record additions to the >$100k ACV customer cohort. Friday’s share-price reaction was dramatically stronger than most of Thursday night’s technology reporters. That supports a useful software taxonomy: AI may pressure software whose primary monetisation unit is a human seat, but it structurally increases search queries, logs, machine-generated events, vector retrieval, security data and observability telemetry. Elastic therefore sits closer to infrastructure consumption than traditional application SaaS. The bull case extends directly to Datadog, and more selectively to Snowflake, MongoDB and Cloudflare: production AI should increase the volume and complexity of data that needs to be stored, searched, monitored and secured. The bear argument remains competition — Amazon Web Services/OpenSearch, Datadog, Cisco/Splunk and Snowflake all overlap portions of Elastic’s addressable market — but Friday’s valuation response tells us the market increasingly views machine-data inflation as one of the safest software manifestations of AI.
3. Workday’s Friday rally reinforces the emerging distinction between “software seats” and “systems of record”: investors are beginning to believe that trusted enterprise context survives the agent transition.
Workday reported Q2 revenue of $2.649bn, +12.8% yoy, subscription revenue +13.9% to $2.471bn and non-GAAP operating margin of 31.1%. More interestingly, management said that more than half of net-new wins included one or more AI products, while AI was cited as a major modernisation driver; shares ultimately rose materially on Friday despite an initially mixed after-hours reaction. The investor debate is shifting away from whether agents eliminate HR/finance seats and towards whether an agent can actually operate enterprise payroll, accounting, permissions and workflow without accessing the authoritative Workday data model. The bull case is that agents make the system of record more important because autonomous actions require clean context, entitlements, audit trails and deterministic business logic. The bear case is that 12–14% growth still looks mature and AI may improve win rates without reaccelerating Workday structurally. The broader read-through is nevertheless constructive for ServiceNow, Salesforce, SAP, Microsoft and Oracle, while still offering much less comfort to lightweight horizontal SaaS products whose workflow logic can be recreated above the application layer.
4. Rubrik’s c.10% sell-off despite one of the strongest cybersecurity prints of the quarter tells us cyber’s valuation bar has also moved sharply higher.
Rubrik delivered subscription ARR of $1.66bn, +33% yoy, revenue +38% to $427.3m, net-new subscription ARR +35%, cloud ARR +39% and a 15% FCF margin, while raising all full-year guided metrics. Yet shares sold off after having risen roughly 48% during August and almost tripled since the 2024 IPO. Fundamentally, very little in the print undermines the thesis. Rubrik is successfully combining cyber recovery with an increasingly ambitious agentic-security strategy around monitoring actions, enforcing guardrails and reversing harmful agent behaviour. The equity debate is simply moving from “is this a high-quality growth asset?” to “how much of that quality is already priced?”. This matters for Palo Alto Networks, CrowdStrike, CyberArk, Zscaler and Cloudflare after powerful AI-security reratings: good results may no longer automatically produce higher multiples. For Commvault, Rubrik’s fundamental strength remains a positive category read-through — cyber resilience appears to be benefiting structurally from AI and ransomware even if the equity reaction becomes less forgiving.
5. SentinelOne exposed the opposite side of the same cybersecurity trade: revenue and ARR were fine, but investors punished even a modest deterioration in the profit trajectory.
Q2 revenue increased 21% yoy to $292m, ARR rose 22% to $1.218bn and non-GAAP operating margin reached a record 10% versus 2% a year earlier. SentinelOne raised its full-year revenue and operating-income outlook, but a reduction in its adjusted EPS outlook helped push shares down roughly 8% on Friday, with CrowdStrike, Zscaler and Okta also weakening in sympathy. This is important because cybersecurity has spent much of 2026 trading on the argument that AI expands TAM through agents, machine identities and autonomous attacks. Investors are now demanding evidence that the incremental opportunity can be captured without delaying margin expansion. The bull case remains that SentinelOne’s endpoint footprint, Purple AI and runtime architecture provide natural distribution for AI security; the bear case is increasingly relative rather than absolute — CrowdStrike and Palo Alto Networks have larger platforms, more telemetry and stronger enterprise consolidation economics. The second-order message for the sector is therefore healthy but demanding: AI-security TAM expansion is real, but “AI” is no longer sufficient to excuse weaker per-share earnings leverage.
6. OpenAI terminating Cursor’s model agreement after SpaceX acquired Anysphere introduces a genuinely new risk to the AI application layer: frontier-model access is not necessarily neutral infrastructure.
Reuters reports that OpenAI will terminate its agreement with Cursor following SpaceX’s reported $60bn acquisition of Anysphere, with access potentially ending from 12 November 2026 under a change-of-control provision. Cursor currently supports models from multiple providers, including OpenAI, Anthropic and Google, alongside models linked to Elon Musk’s broader ecosystem. This matters far beyond the Musk–Altman dispute. A substantial part of the AI application market assumes developers can arbitrage frontier models based on price, latency and quality. If model suppliers increasingly use access as a strategic weapon, application vendors need genuine multi-model architecture and bargaining power, rather than merely exposing a model-selection dropdown. The bull case for Cursor is that the episode proves the value of model neutrality and could strengthen Anthropic and Google relationships. The bear case is that frontier labs can vertically integrate into coding and selectively disadvantage applications that compete with them or align with rivals. This is particularly relevant for GitHub/Microsoft, Atlassian, GitLab, JFrog and emerging private coding platforms: model dependency may become a strategic-risk factor alongside compute costs and customer acquisition.
7. Anthropic’s $45bn Nscale agreement is increasingly becoming the focal point for the neocloud debate: the demand is unquestionably enormous, but the financing architecture is becoming increasingly circular.
Anthropic has reportedly committed $45bn over six years for approximately 460MW of Nvidia Vera Rubin capacity at Nscale’s West Virginia campus. The Financial Times notes that Nscale’s broader development is planned around a 1.35GW data centre plus a 2GW gas facility, while Anthropic’s total infrastructure commitments across providers reportedly exceed $150bn. Today’s more interesting development is the investor debate around the structure: neoclouds frequently lease facilities, borrow against long-term customer contracts, source GPUs from suppliers that are also investors, and serve frontier laboratories themselves dependent on continual external funding. The bull view is that scarcity makes this rational — demand exceeds available powered capacity and contractual commitments create high visibility. The bear view is that rapidly depreciating hardware plus termination clauses can turn apparent long-duration backlog into leveraged duration risk surprisingly quickly. CoreWeave and Nebius remain the clearest public-market read-through; Nscale, Crusoe and other private neoclouds sit on the same continuum. Nvidia also matters because the more capital it commits to customers and infrastructure partners, the harder investors must work to separate organic end-demand from ecosystem-supported demand.
8. SK Hynix is telling investors explicitly that HBM scarcity could persist through 2030, which materially extends the semiconductor-capex cycle beyond current Nvidia generations.
SK Hynix broke ground on its >$4bn Indiana advanced-packaging facility, with US production of next-generation HBM targeted from H2 2029. Separately, management has said it expects memory shortages to persist through 2030, while the company has approved roughly KRW54.3tn ($38.3bn) of investment through 2031. The bull thesis for SK Hynix, Micron Technology and Samsung is increasingly based not just on unit growth but on rising HBM content per accelerator and much greater packaging complexity. Importantly, new capacity arriving in 2029 is too late to solve the next several years of tightness. The bear case is the classic memory-cycle paradox: persistent scarcity stimulates enormous investment, which eventually destroys pricing. But if AI compute grows anywhere near hyperscaler and frontier-lab plans, HBM bit demand could compound quickly enough to absorb considerably more capacity than previous memory cycles. The cleaner second-order exposure may therefore remain Applied Materials, Lam Research, KLA, ASML, Tokyo Electron, Advantest and BE Semiconductor, which monetise the capex irrespective of where eventual HBM pricing settles.
9. Andreessen Horowitz creating a dedicated $1.1bn “Machine Age” hardware fund is a notable capital-allocation signal: Silicon Valley’s investment hierarchy is moving from software abundance towards physical bottlenecks.
The Wall Street Journal reports that the new fund will invest across processors, memory, networking, robotics and other physical AI infrastructure, part of roughly $15bn that Andreessen Horowitz raised earlier this year. Venture investment in semiconductor and autonomous-machine companies has reportedly reached roughly $100bn over the past year. This is not a direct earnings catalyst, but it crystallises an important valuation debate. The previous software era rewarded businesses where marginal distribution cost approached zero; the AI era currently rewards whoever controls scarce compute, power, networking, memory, packaging and physical deployment capability. That capital migration supports the long-duration thesis for Nvidia, Broadcom, Marvell Technology, Arista Networks, Vertiv, Eaton and semiconductor equipment, while it also accelerates competition by funding challengers such as Cerebras and specialised inference silicon. The second-order question is whether this is merely rational capital chasing genuine bottlenecks or the beginning of the next overcapacity cycle. Both can be true sequentially.
10. The broader software tape has now delivered the most meaningful challenge yet to 2026’s “AI eats software” trade — but the rebound is becoming highly selective rather than indiscriminate.
The State Street software/services basket jumped roughly 5.2% on Thursday to a record, driven by strong Salesforce and cybersecurity earnings; Elastic then extended the rebound on Friday, while Workday also finished higher. Microsoft has meanwhile recorded its longest winning streak in roughly ten months. The important change is psychological: six months ago the market increasingly treated AI as a terminal-value threat to most SaaS; this week’s results from Salesforce, CrowdStrike, Okta, Workday and Elastic demonstrate several credible pathways to monetisation or workload expansion. But Marvell, Rubrik and SentinelOne simultaneously show that investors are not simply reverting to the old “beat and raise = stock up” regime. The new hierarchy looks increasingly clear: systems of record, proprietary enterprise context, machine-data infrastructure, cybersecurity enforcement and scarce physical infrastructure are being rewarded; generic workflow/UI layers remain vulnerable. That is a healthier and ultimately more investable framework than either “AI kills all software” or “AI lifts everything”.
Bottom line
The most useful read-through this morning is a change in the market’s hurdle rate. Nvidia’s extraordinary growth outlook has effectively reset expectations for the entire AI complex. Consequently, Marvell Technology can grow 37% and still fall >8%; Rubrik can grow ARR 33% and still fall roughly 10%; SentinelOne can raise revenue guidance and still be punished. Investors increasingly want near-term conversion of AI narrative into revenue, margins and cash flow, not distant TAM.
Conversely, Elastic and Workday were rewarded precisely because their results provide evidence for two of the more durable AI-software control points: machine-data infrastructure and authoritative enterprise systems. In hardware, SK Hynix reinforces the duration of physical scarcity; Anthropic/Nscale reinforces the extraordinary scale of future compute demand but simultaneously keeps financing/circularity risk at the centre of the debate. And OpenAI–Cursor introduces another underappreciated strategic variable: frontier models themselves are becoming competitive distribution platforms, not neutral utilities.
For positioning, I would currently place the highest-quality structural exposures around Nvidia/Broadcom for compute architecture; Arista Networks/Marvell Technology/Credo for connectivity; SK Hynix/Micron Technology plus semiconductor equipment for memory intensity; Elastic/Datadog for machine-data growth; Workday/ServiceNow/Salesforce for system-of-record durability; and Palo Alto Networks/CrowdStrike/CyberArk/Rubrik for security enforcement and governance. The key caveat is valuation: this week demonstrated that even excellent fundamentals can be insufficient once too much future AI success has already been capitalised.
