1. CoreWeave’s Q2 is the strongest overnight confirmation that AI infrastructure demand is still accelerating, but it also intensifies the debate over whether neocloud economics can ever resemble software economics.
Q2 revenue reached $2.58bn, ahead of expectations, while adjusted loss per share narrowed to $1.03 versus c.$1.20 expected. More importantly, backlog increased to $104.2bn from $99.4bn in Q1, with more than $25bn of new customer commitments added this quarter; CoreWeave also raised FY26 capex guidance to $35–39bn from $31–35bn and lifted its revenue and adjusted operating-profit outlook. Shares rose more than 14% after hours. The bull case is that this finally demonstrates operating leverage alongside extraordinary demand visibility: the customer mix is broadening beyond OpenAI/Microsoft into Meta, Anthropic and enterprise accounts, while new contracted capacity effectively underwrites several years of build-out. The bear case is that investors are celebrating a business spending $9.4bn of capex in a single quarter to generate $2.6bn of revenue, with financing, depreciation and power costs structurally unlike traditional cloud software. The second-order implication is unambiguously positive near term for NVDA, VRT, ANET, AVGO, MU and data-centre power/infrastructure suppliers, but strategically more mixed for CRWV itself: the bigger the backlog becomes, the more valuation ultimately depends on return on deployed capital rather than revenue growth. This remains the cleanest test case for whether AI scarcity economics accrue to capacity owners or primarily to Nvidia and the physical suppliers beneath them.
2. Super Micro’s FY27 guidance is arguably an even cleaner physical-demand signal: the AI server market is not just growing—it is broadening materially across customers.
Super Micro guided FY27 revenue to $65–72bn, dramatically above the $52.5bn consensus, sending the shares roughly 7% higher after hours. Q4 revenue nearly doubled to $11.12bn, albeit slightly below consensus because power, cooling and networking delays pushed deployments into the following quarter, while gross margin recovered to 17.5%. Notably, Super Micro had nine customers each generating more than $1bn of annual revenue in FY26, versus four a year earlier. The bull conclusion is important: AI infrastructure is no longer simply Microsoft, Meta and OpenAI buying Nvidia systems; the customer base is becoming sufficiently broad that system vendors can contemplate another year of very high growth even from an enormous revenue base. Bears will focus on the opposite message embedded in the delays—GPU availability is no longer the only constraint; power, cooling and networking increasingly determine shipment timing, which raises execution risk and working-capital intensity. This strengthens the second-order case for VRT, ETN, ANET, AVGO and optical/networking suppliers, while reducing the attractiveness of treating server assemblers as equivalent to Nvidia: systems volumes can explode while gross margins remain in the teens. The value chain continues to reward architectural scarcity more than assembly scale.
3. Foxconn’s Q2 provides independent confirmation from the world’s largest electronics manufacturer that AI servers are becoming a structurally larger profit pool than consumer electronics.
Q2 net profit rose 35% yoy to T$59.97bn (c.$1.86bn), ahead of the T$58.8bn consensus, after revenue increased roughly 40% yoy; Foxconn reiterated expectations for strong FY26 revenue growth and continues expanding AI-server manufacturing capacity in Mexico and Texas for Nvidia-related systems. The strategically important point is that cloud/networking and AI-server activity is rapidly changing Foxconn’s revenue mix away from its historical dependence on Apple hardware. Bulls will see this as another physical shipment datapoint corroborating the enormous capex numbers disclosed by hyperscalers and neoclouds. Bears will argue that Foxconn, Super Micro and CoreWeave are all simultaneously scaling infrastructure against the same end-market demand signal, which increases the probability of overbuild once power constraints ease. For NVDA/AVGO/TSMC, the read-through remains positive because system demand continues translating into real deployments; for AAPL, it reinforces the relative scarcity of components and manufacturing attention being diverted towards AI. The broader debate is shifting from “is the AI capex real?”—the evidence increasingly says yes—to how long can deployment growth remain above underlying AI revenue growth before utilisation becomes the binding KPI?
4. The reported Taiwan breach is potentially the most consequential cyber development this week because it suggests autonomous offensive AI may be moving from controlled research into real-world state-linked operations.
The FT reports that suspected China-linked attackers used open-source AI agents in an attack on Taiwanese government infrastructure in July, with a tool coordinating eight agents to map systems, exfiltrate personnel records and penetrate government and energy-related targets; cybersecurity company Dream described it as the first known “end-to-end autonomous” attack against a government entity, although attribution and some details remain based on Dream’s forensic assessment rather than public confirmation from Taiwanese authorities. The investor significance is larger than another ransomware headline: if open-source agents can autonomously coordinate reconnaissance, exploitation and exfiltration, the attacker cost curve collapses while attack frequency and parallelism rise. That strengthens the structural case for PANW, CRWD, ZS, CYBR, OKTA and MSFT, but especially platforms capable of automated enforcement rather than merely generating alerts. PANW’s network/cloud/runtime positioning and CRWD’s endpoint/telemetry scale look increasingly relevant because humans cannot manually investigate attack volume that itself becomes machine-generated. The bear case is not weaker cyber demand; it is that AI commoditises portions of detection and analysis, concentrating economics further into vendors controlling identity, telemetry and enforcement. This is exactly the environment in which point tools can lose share even as the total cyber TAM expands.
5. Meta’s new small open-weight model sharpens the software debate: frontier intelligence may matter less economically than the proliferation of cheap specialised agents running everywhere.
Meta has launched Muse Glimmer, an open-weight model intended for smaller agentic tasks that can run on a single graphics card, while Zuckerberg has argued for fewer US restrictions on open-weight AI as competition with Chinese models intensifies; Meta also plans a forthcoming higher-end Muse Spark 1.2 release. This matters because the next phase of disruption may not require every enterprise task to call an expensive frontier model. If capable smaller models can run locally or at very low inference cost, agents can proliferate across endpoints, applications and internal workflows, making AI dramatically cheaper to deploy. That is negative for parts of traditional SaaS where simple workflow execution is the product, because customers gain a lower-cost automation substitute; it is also a potential pricing headwind for proprietary frontier-model providers. Yet it is structurally positive for infrastructure and security layers because more agents mean more machine identities, API traffic, telemetry and policy decisions. The second-order beneficiaries therefore remain NET, PANW, CRWD, ZS, CYBR, OKTA and potentially DDOG, while conventional horizontal SaaS must increasingly prove that proprietary data and workflow control—not simply embedded AI—protects pricing power.
Bottom line
this morning materially strengthens the AI infrastructure volume thesis but simultaneously raises the capital-efficiency question. CoreWeave, Super Micro and Foxconn all point to exceptional physical deployment demand; the debate is now how much of that growth ultimately earns attractive ROIC outside NVDA/AVGO and scarce infrastructure control points. In software, the more important structural shift is that cheap autonomous agents appear increasingly capable on both sides of the cyber equation. That remains supportive of cyber and machine-traffic infrastructure while keeping pressure on conventional seat-based SaaS. My preferred exposure hierarchy remains NVDA/AVGO/ANET/VRT in infrastructure and PANW/CRWD/CYBR/ZS in cyber, with greater caution around capital-intensive capacity owners and undifferentiated horizontal SaaS.
