The incremental message this morning is that the AI investment debate is moving beyond GPUs into power ownership, financing structure, cyber-systemic risk and the durability of software control points. Two genuinely new weekend developments matter most: OpenAI appears to have secured substantial economic exposure to one of its power-infrastructure suppliers, while the Financial Stability Board has elevated AI-driven cyber risk into a global financial-stability issue. Against that backdrop, last week’s earnings continue to provide the clearest fundamental read-through: Nvidia says the infrastructure cycle is still accelerating, while CrowdStrike, Salesforce, Workday, Elastic and Rubrik increasingly show that parts of software are monetising AI rather than merely defending against it.
1. OpenAI’s reported $5.5bn SB Energy warrants make the circularity debate substantially more important
The Wall Street Journal, citing draft IPO documents, reported that OpenAI was issued warrants in power-infrastructure company SB Energy valued at roughly $5.5bn. Reuters could not independently verify the report and neither company had commented when Reuters published, so this should be treated as reported rather than company-confirmed.
Strategically, however, this is highly significant. AI labs are no longer merely buying GPU hours; they are becoming economically intertwined with power developers, data-centre operators, neoclouds, semiconductor suppliers and financiers. The bull argument is that OpenAI is locking up scarce electrical infrastructure while simultaneously participating in the upside created by its own demand. The bear argument is that these cross-holdings increasingly muddy the distinction between independent end-demand and ecosystem-supported demand.
For investors, the key point is that power is becoming an investable AI control point in its own right. The beneficiaries extend beyond Nvidia into Vertiv, Eaton, data-centre developers, utilities and private power-infrastructure companies. But it also raises an accounting and quality-of-demand question analogous to vendor financing: when the customer, supplier, investor and warrant holder become increasingly connected, headline backlog deserves more scrutiny.
2. The Financial Stability Board has now elevated AI-driven cyber risk from a technology concern to a systemic-financial-risk concern
Financial Stability Board chair and Bank of England Governor Andrew Bailey warned G20 finance ministers and central-bank governors that rapidly advancing AI could increase the speed, scale and cost-effectiveness of cyberattacks, while financial institutions’ dependence on a small number of technology providers could amplify systemic risk. The warning explicitly referenced recent frontier-model incidents and called for stronger resilience and deployment protocols.
This is materially more important for cybersecurity than another vendor marketing announcement because it brings regulatory compulsion into the AI-security TAM argument. The likely spending implications extend well beyond endpoint detection:
- Identity and machine identity: Okta, CyberArk
- SOC automation and telemetry: CrowdStrike, Palo Alto Networks
- Network and runtime enforcement: Palo Alto Networks, Zscaler, Cloudflare
- Cyber recovery: Rubrik, Commvault
The bull case is that AI security becomes non-discretionary infrastructure. The bear case is that regulatory scrutiny increasingly focuses on systemic dependence on a handful of technology platforms, potentially constraining hyperscaler and security concentration. But for aggregate cyber spend, this looks structurally positive.
3. Nvidia’s Q2 numbers remain the single strongest fundamental datapoint for the entire AI infrastructure cycle
Nvidia reported Q2 FY27 revenue of $96.2bn, +106% yoy, including data-centre revenue of $89.0bn, +117% yoy. Revenue increased another 18% sequentially, while non-GAAP operating income rose 124% yoy to $64.0bn and gross margin held at 75%. Management’s central claim was that multiple frontier labs, open-model developers and physical-AI workloads are now scaling simultaneously, rather than infrastructure demand depending on one or two customers.
The investment debate has therefore shifted from “will AI capex collapse?” towards how long 70–100% infrastructure growth can persist before power, financing or custom silicon become the binding constraint. The most important second-order read-throughs remain positive for TSMC, SK Hynix, Micron Technology, Arista Networks, Broadcom, Marvell Technology, Credo, Astera Labs, Coherent, Lumentum, Vertiv and Eaton.
The bear case increasingly centres on capital intensity rather than demand: extraordinarily high infrastructure returns attract extraordinarily high investment. But Nvidia’s Q2 still argues that the industry is capacity-constrained today, not demand-constrained.
4. CrowdStrike delivered perhaps the strongest evidence yet that “AI expands cyber TAM” is converting into ARR
CrowdStrike’s Q2 FY27 was unusually strong: revenue increased 26% yoy to $1.47bn, record net-new ARR reached $333m, +51% yoy, and ending ARR from customers adopting Falcon Flex reached $2.29bn, +101% yoy. Management raised full-year net-new ARR growth guidance by 630bp to 34% at the midpoint, while Q2 free cash flow reached a record $377m.
This matters much more than generic AI-security product launches. CrowdStrike is showing that the current threat environment is producing measurable wallet consolidation and faster ARR creation. The bull case is that endpoint remains the privileged telemetry layer for humans, workloads and increasingly AI agents, allowing CrowdStrike to attach SIEM, identity, cloud, data and agent-security products through Flex. The bear case is valuation and competitive overlap with Palo Alto Networks: both companies increasingly want to become the enterprise’s broad security operating system.
Fundamentally, CrowdStrike’s quarter materially weakens the thesis that frontier models disintermediate cyber vendors. The opposite appears to be happening: frontier AI is increasing the number and speed of objects that need securing.
5. Salesforce’s Q2 is the most important rebuttal yet to the blanket “AI kills enterprise SaaS” argument
Salesforce reported Q2 FY27 revenue of $11.3bn, +11% yoy, cRPO of $33.5bn, +14% yoy in constant currency, non-GAAP operating margin of 34.1%, and FCF of $1.1bn, +81% yoy. Salesforce also raised FY27 revenue guidance. More strategically, Salesforce and Anthropic announced Claudeforce, allowing Claude to act over Salesforce data, workflows, business logic and governance rather than attempting to replace the underlying system.
That architecture matters enormously. The simplistic bear argument was that the agent replaces the application. Salesforce increasingly argues for a different chain: agent, then system-of-record data, then governed action. If that becomes the dominant enterprise architecture, systems possessing proprietary customer data, permissions, workflow logic and audit trails may become more valuable to agents, not less.
That is constructive for Salesforce, ServiceNow, Workday, SAP, Oracle and Microsoft, and less reassuring for lightweight horizontal software where the user interface itself represents most of the moat.
6. Anthropic’s MatX discussions reinforce the custom-silicon threat to merchant inference economics
Reuters reports that Anthropic discussed acquiring AI-chip startup MatX for approximately $7bn before talks shifted towards a potential partnership. MatX was founded by former Google TPU engineers, and Anthropic is separately building out internal chip capability. Combined with Anthropic’s enormous infrastructure commitments, this tells us something important about frontier-model economics: compute cost is now large enough to justify vertical integration into silicon.
This is not necessarily bearish for Nvidia in training. Frontier-model architectures change rapidly, and that is precisely where Nvidia’s programmable stack and CUDA ecosystem are enormously valuable. The more credible long-run segmentation remains:
- Frontier training and rapidly evolving workloads stay Nvidia-heavy
- Large repetitive inference workloads see rising custom-ASIC penetration
That remains structurally positive for Broadcom, Marvell Technology, TSMC, Synopsys, Cadence and Arm, because more companies designing proprietary chips expands the silicon ecosystem even if merchant-GPU share eventually moderates.
7. Marvell Technology confirms custom silicon is becoming a second enormous AI semiconductor market — but its share-price reaction exposed the new valuation hurdle
Marvell reported record Q2 FY27 revenue of $2.739bn, +37% yoy, with data-centre revenue growth accelerating to 46%. Management described AI bookings as “exceptionally robust”, expects a significant acceleration in custom silicon beginning in H2 FY27, and again raised FY27 and FY28 revenue expectations. Fundamentally this is an excellent read-through for the custom-compute architecture.
But the market’s scepticism around the timing of some future hyperscaler ramps shows that AI design wins are no longer sufficient on their own. Investors increasingly distinguish near-term production revenue from large but distant TAM. That is important for Broadcom as well, which currently enjoys greater investor confidence because a larger proportion of its custom-AI opportunity is perceived to be in production or nearer-term ramp. For Credo, Astera Labs, Coherent, Lumentum and Arista Networks, the broader message remains bullish: both custom accelerators and Nvidia systems consume enormous amounts of connectivity.
8. Workday supports the “systems of record survive agents” thesis from a second major enterprise-software vertical
Workday reported Q2 FY27 revenue of $2.649bn, +12.8% yoy, subscription revenue of $2.471bn, +13.9% yoy, and non-GAAP operating margin of 31.1% versus 29.0% a year earlier. The important debate is not whether AI eventually reduces some HR or finance seats — it almost certainly will. The more important question is where the agent gets trusted payroll, employee, accounting, permission and organisational data.
Workday remains the authoritative data layer for many of those actions. Autonomous agents may therefore reduce the importance of the human interface while increasing the importance of the underlying data model and transaction engine. The bull case is durability plus potentially higher workflow penetration; the bear case is that roughly 14% subscription growth still does not demonstrate material AI-driven reacceleration. I would read Workday as constructive for ServiceNow, Salesforce, SAP and Oracle, but not as evidence that all SaaS is safe.
9. Elastic is showing exactly what investors should want from AI-exposed infrastructure software
Elastic reported Q1 FY27 revenue of $478m, +15% yoy, but cRPO increased 21% to $1.153bn, while sales-led subscription revenue grew 18% and additions to customers exceeding $100k ACV reached a record level. Contracted demand is accelerating faster than reported revenue, which is arguably one of the cleanest software expressions of the AI workload thesis:
- Agents create logs
- Inference creates telemetry
- Vector search creates queries
- Machine activity generates security events
Unlike seat-based SaaS, Elastic’s relevant underlying unit — machine-generated data — should expand as AI adoption rises. The strongest read-through is for Datadog, followed by Snowflake, MongoDB and parts of Cloudflare. The bear case remains intense competition from Amazon Web Services, Datadog, Cisco/Splunk and hyperscaler-native services. But if cRPO continues outrunning revenue, the market will increasingly view observability and search as one of the more structurally AI-beneficial software categories.
10. Identity and cyber recovery are emerging as two distinct AI control points
Okta reported Q2 revenue of $805m, +11% yoy, but RPO grew 17%, cRPO 14%, and FCF margin reached 28%. Management explicitly framed autonomous agents as identities that require discovery, authentication, governance and controls over what they can do. Rubrik, meanwhile, reported subscription ARR of $1.66bn, +33% yoy, revenue of $427.3m, +38% yoy, operating cash flow margin of 18% and FCF margin of 15%, while raising all FY27 guided metrics.
These businesses attack different consequences of autonomous AI. Okta answers who or what the agent is, and what it is permitted to access. Rubrik answers what happens when an agent corrupts, deletes or exposes critical data. That suggests AI security is evolving into a much broader architecture than endpoint detection alone.
The strategic winners are likely vendors owning a hard enforcement or recovery primitive — identity, network policy, endpoint telemetry, data protection or privileged access — rather than standalone AI-analysis features that frontier models can increasingly reproduce.
Bottom line
The biggest new development this morning is OpenAI’s reported economic entanglement with SB Energy. It extends the AI circularity debate from GPUs and neocloud contracts directly into power infrastructure. The industry is increasingly becoming a tightly coupled capital system in which frontier labs, semiconductor suppliers, data-centre operators, power companies and financiers all participate economically in each other’s growth.
At the same time, the Financial Stability Board’s intervention materially strengthens the medium-term cybersecurity thesis. AI-driven cyber risk is now being framed as a financial-stability issue, not merely an enterprise IT issue. That should reinforce spending around identity, autonomous SOC, runtime enforcement and cyber recovery.
And last week’s earnings give us a more useful software hierarchy. CrowdStrike and Rubrik show AI increasing security demand; Elastic shows machine-generated data expanding infrastructure workloads; Salesforce and Workday suggest authoritative systems of record retain strategic value in an agentic world. The vulnerable part of software remains the layer where the product is principally a user interface or thin workflow sitting above somebody else’s data and model.
For positioning, the highest-quality structural exposure still clusters around Nvidia, Broadcom and TSMC for compute; SK Hynix and Micron Technology for memory; Arista Networks, Marvell Technology and Credo for connectivity; Vertiv and Eaton for power infrastructure; Datadog and Elastic for machine data; Salesforce, ServiceNow and Workday for enterprise context; and Palo Alto Networks, CrowdStrike, CyberArk and Rubrik for security enforcement and recovery. The growing caveat across the entire AI complex is no longer end-demand; it is valuation, financing quality and how much future success has already been capitalised.
