The most important incremental development this morning is OpenAI’s Jalapeño benchmark disclosure. Until yesterday, custom silicon was largely a hyperscaler strategy and an eventual Nvidia risk; OpenAI has now published measured results from an inference ASIC designed around its own workloads. Combined with Nvidia reporting tonight, Salesforce, CrowdStrike and Okta reporting alongside it, today becomes an unusually clean test of where the economic rents from AI are actually migrating — silicon, infrastructure, application software or security. Markets enter the day somewhat less hostile to duration after oil and Treasury yields eased overnight, with Nvidia recovering 2.2% yesterday and the semiconductor index +1.4%.
1. OpenAI’s Jalapeño benchmarks are potentially the most important competitive datapoint for Nvidia since Google began scaling TPUs: custom inference silicon has moved from strategic aspiration to measured working hardware.
OpenAI disclosed yesterday that Jalapeño, its first custom inference ASIC co-developed with Broadcom, achieved higher throughput per kilowatt and lower token latency than the commercial systems in its comparison across GPT-OSS 120B, DeepSeek R1 and Kimi K2.5. OpenAI rates the chip at 700W, with measured sustained consumption of ≤550W in tested workloads; third-party reporting on the disclosed InferenceX results puts the claimed advantage at roughly 1.5–1.9× throughput per kilowatt and 1.7–3.6× lower end-to-end latency versus selected Nvidia GB200/GB300 configurations. OpenAI plans initial deployment by end-2026 and describes Jalapeño as the first generation of a multi-generation platform that it intends to deploy at gigawatt scale. The caveat matters: these remain OpenAI-selected benchmarks, Jalapeño is inference-only, Rubin was not the comparison point and Nvidia will have advanced its own architecture by the time Jalapeño scales. So I would not read this as “OpenAI has beaten Nvidia”. The important investor conclusion is economic: hyperscalers and frontier labs now have increasingly credible ways to optimise their highest-volume inference workloads around proprietary silicon. That is structurally positive for AVGO, which gets the design/implementation, networking and connectivity economics regardless of who owns the accelerator IP, and for TSMC/HBM/networking/optics. For NVDA, training and frontier workloads remain strongly protected by CUDA and the full-stack system, but the assumption that Nvidia captures a roughly constant share of every incremental inference dollar becomes harder to sustain. The most uncomfortable read-through may again be for AMD: the industry is increasingly bifurcating between Nvidia’s integrated platform and customers’ own ASICs, leaving less obvious strategic space for a generic second merchant accelerator.
2. That makes Nvidia’s earnings tonight more consequential, not less: investors are now simultaneously testing Blackwell/Rubin demand against credible custom-silicon substitution and increasingly visible vendor financing.
Consensus cited by Reuters expects Q2 revenue of roughly $92.2bn, nearly +100% yoy, followed by Q3 revenue around $104.2bn, +82.8%, with gross margin around 75%. Yet the market already knows demand is exceptional. The harder questions are the transition to Rubin, inference mix, gross-margin sustainability and whether Nvidia must increasingly finance the ecosystem consuming its chips. Nvidia has arranged financing initiatives targeting >$500bn of AI infrastructure and guaranteed up to $105bn supporting OpenAI’s Ohio data-centre lease, prompting some investors to describe it increasingly as a “central bank” for the AI ecosystem. The bull case is still formidable: if Nvidia can sustain c.80%+ forward growth while customers absorb rising HBM/server prices and Rubin ramps cleanly, earnings revisions can overwhelm both higher rates and custom-ASIC fears. The bear case has become subtler. Nvidia can beat numbers and still struggle if incremental demand increasingly requires guarantees, equity investments or financing structures, because investors then have to distinguish organic customer ROIC from demand enabled by Nvidia’s balance sheet. OpenAI’s Jalapeño announcement amplifies the second-order issue: the more expensive Nvidia’s systems become — customers have reportedly been warned of >15% server-price increases in early 2027 — the stronger hyperscalers’ economic incentive to shift stable, high-volume inference towards ASICs. My read-through hierarchy tonight is therefore: Rubin demand/gross margin first, financing exposure second, China third. A strong result remains positive for AVGO/ANET/VRT/MU, but custom silicon can simultaneously gain strategic share even in a bullish Nvidia demand environment.
3. Salesforce tonight is the cleanest application-layer test yet of whether AI is genuinely monetising inside incumbent SaaS or merely making existing products more useful.
The Street is looking for roughly $11.3bn of Q2 revenue, around +11% yoy, with investors focused much more heavily on cRPO and Agentforce economics than headline EPS. Salesforce entered the quarter with Agentforce ARR around $1.2bn, but the core debate is whether that activity translates into incremental contracted spend and consumption quickly enough to offset mature seat-based workloads. This distinction has become particularly important as model inference gets cheaper and “vibe coding” makes lightweight internal applications easier to create. The SaaS bear case says cheaper agents reduce the value of both licences and simple workflow software; Salesforce is arguably one of the best places to test that thesis because CRM data, permissions, audit history and process integration should be unusually difficult to recreate. The bull case is therefore not simply “Agentforce ARR grows”: investors need core Sales/Service resilience plus Agentforce/Data Cloud increasing revenue per customer. If cRPO meaningfully accelerates and management can demonstrate that customers using agents expand rather than shrink their Salesforce footprint, the read-through is powerful for NOW, WDAY, SAP and TEAM, weakening the indiscriminate “AI eats software” thesis. If Agentforce adoption looks impressive but core organic growth remains soft, that would be more damaging: it would suggest incumbents can successfully deploy AI without necessarily capturing the productivity surplus economically. Today’s software print is thus fundamentally about who owns the surplus from AI automation — customer or vendor.
4. CrowdStrike and Okta reporting tonight provide almost the perfect control experiment against Salesforce: cyber should theoretically gain units from AI rather than lose seats, so the question is whether that theoretical advantage is finally visible in ARR.
CrowdStrike ended Q1 with $5.51bn ARR, revenue growth of roughly 26%, a 34% FCF margin and accelerating net-new ARR; importantly, its AIDR product has reportedly seen ARR grow >250% sequentially, with management positioning Falcon explicitly as an “agentic security platform”. Okta, meanwhile, now describes its addressable identity perimeter as AI, machine and human identities and reports tonight as well. The structural difference versus conventional SaaS is critical: every autonomous agent can create another identity, credential, endpoint relationship, API session and privileged action requiring governance even if the agent replaces a human worker. That gives CRWD/OKTA/CYBR/PANW/ZS the potential to benefit from both the offensive and enterprise-adoption sides of AI. But cyber valuations have already started reflecting this asymmetry, so tonight needs numbers rather than narrative. For CRWD I would focus on net-new ARR, Falcon Flex consumption, AIDR attach and whether AI-related products are genuinely incremental; for OKTA, watch whether machine/non-human identity begins to alter growth expectations rather than remaining an architectural talking point. A strong pair of prints would be one of the clearest pieces of evidence yet that cyber deserves a structurally higher relative multiple than application SaaS because AI increases its monetisable unit base. Weak numbers would not invalidate the TAM thesis, but would tell us monetisation is lagging strategic relevance — important ahead of PANW’s 1 September results.
5. Apple’s M6/M5 Ultra launch is easy to dismiss as a PC refresh, but strategically it strengthens the case for a parallel AI architecture built around local agents rather than every inference request going back to the cloud.
Apple yesterday unveiled the first M6 Mac mini alongside M5 Max/M5 Ultra Mac Studios, explicitly targeting users running AI workloads and agents locally. The Mac Studio can support up to 512GB of unified memory, while Apple is positioning the machines at developers and professionals running large models and persistent local AI workflows. This does not threaten hyperscaler AI capex near term — frontier models, training and complex inference remain cloud-intensive — but it creates an important second-order debate. If smaller/open models become sufficiently capable, private and cheap to run locally, AI compute splits across cloud, enterprise edge and endpoint rather than centralising entirely in hyperscale data centres. That is favourable for memory content and advanced-node silicon while potentially lowering the marginal cloud inference demand for certain workloads. More interestingly for software/cyber, local autonomous agents dramatically expand endpoint privilege and data-access risk, strengthening the architectural case for CRWD, PANW, ZS, OKTA/CYBR and potentially Apple’s own security stack. The emerging AI infrastructure model therefore looks less like one giant cloud and more like a hierarchy: frontier training in centralised GPU/ASIC clusters, large-scale inference in custom hyperscaler silicon, and increasingly capable agents running locally at the edge. Different semiconductor and software vendors capture rents at each layer.
Bottom line
today is effectively a four-layer AI referendum. OpenAI has demonstrated that custom inference silicon is becoming technologically credible, sharpening the long-run threat to Nvidia’s inference share while materially improving Broadcom’s strategic position. Nvidia tonight tests whether the infrastructure layer can continue producing extraordinary growth and margins without increasingly extraordinary financial support. Salesforce tests whether incumbent SaaS can capture the economic surplus from agents. CrowdStrike and Okta test whether cybersecurity can convert the explosion in machine activity into measurable ARR. My preferred architecture still favours NVDA/AVGO/ANET/VRT across compute and connectivity and PANW/CRWD/CYBR/ZS across mandatory security control points, but Jalapeño makes me incrementally more constructive on AVGO/custom silicon relative to the assumption of indefinitely stable Nvidia inference share. The single most useful comparison tomorrow morning will not simply be who beat consensus; it will be which layer showed the clearest evidence that AI is improving incremental economics rather than merely increasing activity or capital requirements.
