The most important message this morning is that the AI trade is broadening rather than fading. Nvidia’s results earlier this week removed the near-term demand objection; last night’s prints from Marvell Technology, Workday, Rubrik, SentinelOne, Elastic and Autodesk then showed that the second-order beneficiaries are beginning to produce cleaner evidence across custom silicon, enterprise agents, cyber resilience and data infrastructure. The debate is therefore shifting from “is AI demand real?” towards which layers retain the economics as infrastructure spend proliferates and software monetisation becomes measurable.
1. Marvell Technology is the strongest incremental semiconductor read-through this morning: custom silicon and connectivity are accelerating together, materially strengthening the case that AI compute will support multiple profit pools beyond Nvidia.
Marvell reported Q2 revenue of $2.739bn, +37% yoy, with data-centre revenue +46% yoy to $2.172bn and now representing 79% of group revenue. More importantly, management said AI-related bookings remain “exceptionally robust”, expects a significant acceleration in Custom beginning in H2 FY27, and raised its FY27 and FY28 revenue outlooks. Q3 revenue guidance of $3.15bn ±5% implies another material sequential step-up. The bull argument is increasingly straightforward: hyperscalers want more proprietary accelerators but still need an external ecosystem for SerDes, optical DSPs, switching, interconnect and increasingly ASIC implementation. That structurally expands the addressable market for Marvell Technology and Broadcom even if Nvidia continues compounding rapidly. The bear case is customer concentration and vertical integration — the very hyperscalers driving custom ASIC demand also have the resources eventually to internalise more of the stack — but today’s numbers say that risk is not currently slowing orders. Second-order exposure: Broadcom is the clearest positive read-through; Credo, Astera Labs, Arista Networks, Coherent and Lumentum benefit from the same bandwidth-density problem, while AMD faces the strategic squeeze between Nvidia on one side and hyperscaler-custom silicon on the other.
2. Workday delivered one of the most important application-software datapoints of the quarter: AI is now contributing directly to bookings rather than merely supporting the narrative.
Q2 subscription revenue grew 13.9% yoy to $2.471bn, 12-month subscription backlog grew 14.2% to $9.034bn, and non-GAAP operating margin expanded to 31.1% from 29.0%. Crucially, Workday said AI drove more than 25% of new ACV, while more than 5,500 customers use at least one organic Workday agent, up >35% sequentially. Management raised FY27 non-GAAP operating-margin guidance to 31.0% while maintaining roughly 13% subscription growth. This matters because it is harder to dismiss than generic “AI engagement” commentary: a quarter of new ACV touching AI suggests enterprise agents can become a monetisation mechanism rather than simply a bundled feature. The bull case is that Workday’s HR/finance system-of-record data, permissions and deterministic workflows create precisely the trusted rails that agents need; the bear case is that 13% underlying subscription growth remains mature and AI may initially support deal conversion without materially reaccelerating the company. Second-order read-through: positive for ServiceNow, Salesforce and SAP, and more selectively for Oracle and Microsoft. It is less reassuring for lightweight horizontal SaaS where an AI agent can reproduce much of the workflow without needing a proprietary system of record. The software debate continues migrating from seat counts towards ownership of enterprise context, data and agent execution rights.
3. Rubrik’s Q2 is arguably the cleanest evidence yet that AI can expand an existing cybersecurity category rather than simply create another product module.
Subscription ARR increased 33% yoy to $1.66bn, net-new subscription ARR increased 35%, cloud ARR increased 39% to $1.48bn, and revenue grew 38% to $427.3m. At the same time, subscription ARR contribution margin improved to 14.0% from 9.4%, free cash flow margin reached 15%, and Rubrik raised every FY27 guided metric, including subscription ARR to $1.880–1.885bn and revenue to $1.685–1.693bn. The investor debate is whether Rubrik remains principally a high-growth cyber-recovery platform or becomes a broader data + agent-security control point. Its Agent Cloud strategy — monitoring agent actions, enforcing guardrails and potentially undoing harmful agent behaviour — is strategically significant because autonomous agents create a new class of operational mistakes that conventional endpoint detection cannot necessarily reverse. The bear argument is product adjacency: agent security is becoming crowded very quickly, with Palo Alto Networks, CrowdStrike, CyberArk, SentinelOne and hyperscalers all targeting related control points. But Rubrik’s numbers suggest the core business is strong enough to fund that expansion without asking investors to underwrite a speculative pivot. Read-through: positive for cyber resilience broadly, especially Commvault, while strengthening the case that data security will remain one of the highest-growth cybersecurity subsegments.
4. SentinelOne’s quarter reinforces the same AI-security thesis from a different architectural direction: endpoint is becoming an AI-runtime and telemetry control plane, and profitability is improving quickly enough that the story is no longer purely about growth.
Revenue grew 21% yoy to $292m, ARR 22% to $1.218bn, and non-GAAP operating margin reached 10% versus 2% a year ago; the company raised FY27 revenue guidance to $1.202–1.207bn and non-GAAP operating income to $124–128m. The interesting debate is not whether SentinelOne can talk credibly about “AI for Security” and “Security for AI” — virtually every cyber vendor can — but whether Purple AI, Prompt Security, cloud/data security and Flex-style platform consumption materially change ARR growth. Management commentary indicates those emerging AI-security products are scaling quickly, while RPO growth has reportedly accelerated sharply. The bull case is that SentinelOne can leverage its endpoint footprint into agent/runtime security in a manner similar to CrowdStrike; the bear case remains relative competitive scale and the risk that CrowdStrike and Palo Alto Networks absorb most incremental enterprise wallet through larger platform contracts. Second-order implication: this is another datapoint against cyber disintermediation. AI appears to be increasing security objects and product attach faster than it reduces conventional endpoint relevance.
5. Elastic quietly delivered one of the better infrastructure-software results of the week, and the key datapoint is that contracted demand is accelerating faster than reported revenue.
Revenue grew 15% yoy to $478m, but cRPO grew 21% to $1.153bn, total RPO grew 27% to $1.854bn, sales-led subscription revenue grew 18%, and the number of customers above $100k ACV rose to >1,800 from >1,550 a year ago. Elastic also generated $143m adjusted FCF and guided FY27 to roughly $2.0bn revenue, 19.4% non-GAAP operating margin and 21.5% adjusted FCF margin. The investor debate here cuts directly into the observability/search/data architecture thesis. AI applications create vastly more logs, vector data, retrieval workloads and machine-generated telemetry; Elastic can monetise that regardless of which model wins. The bull case is therefore that search + observability + security convergence makes Elastic a classic “picks-and-shovels” software beneficiary. The bear case is competitive intensity from Datadog, Splunk/Cisco, OpenSearch/Amazon Web Services, Snowflake and hyperscaler-native tooling. The fact that RPO is materially outrunning revenue is nevertheless a useful signal that demand is not merely usage volatility. Read-through: positive for Datadog and the broader observability/data layer, but it also raises the bar: companies in this bucket should increasingly be capable of showing contracted acceleration as AI moves into production.
6. Autodesk’s results strengthen the emerging argument that vertical software may be materially more AI-defensible than generic horizontal SaaS.
Q2 revenue increased 16% yoy to $2.046bn, non-GAAP operating margin reached 41%, FCF increased 24% to $561m, and Autodesk raised FY27 revenue guidance to $8.295–8.345bn, with FCF now expected at $2.725–2.750bn. The investment thesis is less about whether Autodesk has a chatbot and more about whether AI can replicate decades of proprietary design context, engineering constraints, file formats, collaborative workflows and compliance requirements. Management’s framing — that useful AI in the physical world requires trusted data and real-world context — is strategically important. The bull case is that vertical SaaS vendors such as Autodesk, Dassault Systèmes, Bentley Systems, PTC and Nemetschek can turn AI into higher workflow penetration rather than seat destruction. The bear case is that generative design eventually lowers the technical barrier to creating and modifying assets, allowing new interfaces to sit above incumbent CAD systems. Near term, however, Autodesk is showing exactly the financial profile investors want from an AI-defensible incumbent: double-digit growth plus margin expansion plus higher cash generation.
7. Anthropic is now seriously exploring proprietary silicon, which makes custom compute a frontier-model strategy rather than merely a hyperscaler strategy.
Reuters reports that Anthropic explored acquiring AI-chip start-up MatX for roughly $7bn before the talks shifted towards a potential partnership; MatX, founded by former Google TPU engineers, is reportedly seeking $4bn of funding. Anthropic has meanwhile committed enormous sums to third-party compute, including a reported $45bn six-year Nscale agreement tied to a 460MW West Virginia site using Nvidia Vera Rubin systems. These are reported negotiations rather than confirmed Anthropic disclosures and should be treated accordingly. The strategic message is more important than the transaction status: once frontier labs are spending tens of billions annually on inference and training, silicon becomes too large a cost pool to leave entirely to merchant vendors. OpenAI’s Broadcom-linked Jalapeño ASIC and Anthropic’s MatX exploration point in the same direction. The bull case for Nvidia is that training complexity, CUDA and system-level integration keep the hardest workloads on its stack even as customers experiment; the bear case is that stable, high-volume inference gets systematically peeled away into custom ASICs. The most obvious beneficiaries are Broadcom, Marvell Technology, TSMC, Synopsys, Cadence, Arm, HBM suppliers and advanced packaging — essentially everyone that earns money when customers design more chips, regardless of who wins accelerator share.
8. SK Hynix is effectively telling investors that the HBM shortage could last to 2030 — which meaningfully extends the memory-cycle duration, but also raises the eventual overbuild risk.
The company broke ground on its $4bn Indiana HBM packaging and R&D facility, targeting HBM4E mass production there from Q3 2029, while management said it sees memory shortages persisting through 2030. SK Hynix has also approved roughly KRW54.3tn ($38.3bn) of investments through 2031. The bull debate is obvious: every increase in accelerator compute density requires disproportionately more HBM bandwidth and content, and the qualification barriers make supply slower to respond than conventional DRAM. That is bullish for SK Hynix, Micron Technology and, selectively, Samsung. The bear case is exactly what makes memory historically dangerous: today’s scarcity creates tomorrow’s capacity. The crucial distinction is whether AI demand compounds quickly enough for HBM content growth to absorb a multi-year capacity build without collapsing pricing. Second-order winners: semiconductor equipment and advanced packaging — ASML, Applied Materials, Lam Research, KLA, Tokyo Electron, Advantest and BE Semiconductor — whose economics depend more on wafer and packaging intensity than on the eventual memory price.
9. More than 100 technology companies are now publicly framing advanced AI cyber capability as an imminent systemic threat, which moves “AI increases cyber TAM” from vendor marketing towards an ecosystem-wide operating assumption.
A coalition including OpenAI, Anthropic, Microsoft, Alphabet, Amazon Web Services and CrowdStrike called for co-ordinated defensive action against AI-enabled cyberattacks, warning that the industry has a limited window to strengthen defences. The investor debate is increasingly about the shape of cyber spend rather than whether spend rises. If frontier models dramatically increase attacker productivity, enterprises will favour security platforms that can operate at machine speed: telemetry, identity, runtime enforcement, automated SOC, recovery and agent governance. That supports Palo Alto Networks, CrowdStrike, Zscaler, CyberArk, Rubrik, SentinelOne and potentially Cloudflare. The bear case for individual equities is consolidation: a larger cyber TAM does not imply every vendor wins. In fact, AI-driven operational complexity could accelerate platform consolidation because customers cannot realistically stitch dozens of point tools into autonomous workflows. Second-order read-through: the strongest beneficiaries should be vendors owning enforcement or proprietary telemetry, rather than standalone analysis layers that an LLM can increasingly reproduce.
10. The Anthropic–Pentagon ruling is strategically important because frontier AI governance is becoming a commercial moat and a commercial liability at the same time.
A US federal judge on 28 August blocked the Pentagon’s move to blacklist Anthropic as a national-security supply-chain risk, calling the measures unlawful; the dispute arose from Anthropic’s restrictions around certain military surveillance and autonomous-weapons uses of Claude. The immediate financial implication is removal, at least temporarily, of a potentially material federal-contracting overhang. The broader investor debate is more consequential: frontier labs are no longer just software companies competing on benchmarks — they are increasingly forced to make sovereign, military and national-security policy choices that can affect addressable markets. OpenAI has generally taken a more permissive government-partnership approach; Anthropic has positioned itself around stronger usage boundaries. The bull case is that trust, controllability and governance become premium enterprise features, particularly in regulated sectors. The bear case is that restrictions push government and sensitive workloads towards rivals willing to permit broader deployment. For cybersecurity, the second-order implication is positive either way: as model providers disagree about what agents should be allowed to do, enterprises increasingly need an independent policy/enforcement layer around those agents.
What matters most
The strongest combined signal from the last 48 hours is that the AI value chain is becoming more distributed, but not less valuable. Nvidia remains the dominant merchant-compute platform, yet Marvell Technology is showing accelerating custom silicon and connectivity demand; Anthropic and OpenAI are independently moving towards proprietary compute; SK Hynix sees HBM scarcity lasting years; and infrastructure-software vendors are beginning to show RPO and ARR acceleration as AI workloads enter production.
On software, the hierarchy is also becoming clearer. Workday and Autodesk support the “systems of record / vertical context survive AI” thesis. Elastic supports the “machine data explodes” thesis. Rubrik and SentinelOne support the “AI expands cybersecurity units” thesis. That is materially different from saying all software benefits from AI.
My relative read-through this morning is therefore: Broadcom and Marvell Technology remain the most interesting non-Nvidia ways to express custom compute; SK Hynix/Micron Technology + advanced packaging/equipment remain beneficiaries of physical bottlenecks; Workday/ServiceNow/Salesforce are increasingly credible AI-monetisation incumbents where proprietary enterprise context matters; and within cyber, Palo Alto Networks, CrowdStrike, Rubrik, CyberArk and Zscaler remain the best-positioned control-point exposures as agent proliferation moves from narrative to production.
