Morning View
The market is beginning to separate AI demand from AI economics. Qualcomm’s new Amazon collaboration is the clearest fresh example: custom inference silicon and 1.6T optical connectivity are moving beyond the Broadcom/Marvell duopoly, but the associated warrant structure shows how commercial commitments and strategic equity incentives are becoming intertwined. At the same time, ServiceTitan and Braze both delivered better-than-expected quarterly growth and materially stronger cash generation, yet their shares sold off after hours as investors focused on forward growth, margin quality and execution risk. The message is not that software demand is collapsing; it is that AI exposure no longer excuses a weak expectations setup.
The cross-sector read-through is mixed but constructive for the physical AI stack. ASML is addressing a technical limitation that has constrained High-NA EUV adoption for very large AI dies, while Boston Scientific’s cyberattack demonstrates that digital-security failures can now directly interrupt manufacturing, shipments and full-year earnings. Conversely, credit markets are applying tighter terms to AI infrastructure projects even as AI-linked debt issuance approaches unprecedented scale, which should favor hyperscalers and other balance-sheet-rich buyers over leveraged developers and neoclouds.
I cross-checked the latest earnings calendar and corporate investor-relations releases before ranking today’s list. ServiceTitan and Braze reported after the September 8 US close and are included below; no other material large or strategically important technology result had been released by the cutoff. SailPoint is scheduled to report fiscal Q2 2027 before the US market opens today, while Apple’s product event is later today and Oracle and Adobe report on September 10.
1. Qualcomm–Amazon turns custom inference into a broader competitive market — but the $60bn figure is a warrant threshold, not backlog
Qualcomm and Amazon announced a multi-generation collaboration covering customized AI data-center silicon and optical connectivity, including links up to 1.6T. The most important disclosure sits in Qualcomm’s SEC filing rather than the headline: Amazon received a warrant for up to 25m Qualcomm shares at $161.26, with vesting tied to commercial arrangements, binding purchase orders and actual Amazon purchases. The warrant mechanics reference up to $60bn of aggregate payments, but that is a maximum commercial-payment threshold for vesting, not a confirmed $60bn order or revenue commitment. About 3.75m shares vested at issuance based on initial purchase commitments. Qualcomm shares finished roughly 3% higher on September 8.
For investors, the deal changes two narratives. First, Qualcomm’s data-center ambition is becoming more credible just as Apple modem insourcing threatens a legacy smartphone profit pool; management is targeting $15bn of data-center revenue by 2029. Second, Amazon is signaling that custom inference will be multi-vendor even though it already develops Trainium and Inferentia internally. The bull case is that Qualcomm can combine Arm-based compute, inference accelerators and Alphawave-derived connectivity into a new growth engine. The bear case is that customer-specific silicon is concentrated, capital intensive and exposed to Amazon’s bargaining power, while Nvidia, Broadcom and Marvell remain formidable. TSMC, Arm, Synopsys, Cadence and optical/networking suppliers benefit from greater design diversity. The next datapoint is how quickly initial purchase commitments convert into recognized revenue and whether Qualcomm names additional hyperscale customers.
Sources: Qualcomm announcement | Qualcomm SEC filing | Reuters
2. ServiceTitan beats the quarter, but a light Q3 guide and CRO transition expose how little tolerance remains for vertical-software deceleration
ServiceTitan reported fiscal Q2 2027 revenue of $292.8m, up 21% year over year and above its own $284–286m guidance range, while non-GAAP operating income reached $44.4m and non-GAAP free cash flow exceeded $50m. Gross transaction volume rose 17% to $26.8bn, net dollar retention remained above 110%, and adoption of its AI-oriented Max offering exceeded the company’s internal target; ServiceTitan now expects more than 700 enrolled Max locations by fiscal year-end. The tension is in the forward setup: Q3 revenue guidance of $285–287m was slightly below available consensus at the midpoint, and longtime Chief Revenue Officer Ross Biestman will step back after Q3, with internal veteran Rikus Pretorius taking over in fiscal Q4. Shares fell more than 19% after hours.
The fundamental change is modestly positive, not 19% negative: the quarter itself showed improving operating leverage and cash conversion, and the CRO succession is internal and planned. But the market reaction is a useful valuation signal. ServiceTitan has been treated as AI-defensible vertical software because it owns trade-specific workflow, payments and operational data; investors now want proof that this moat produces sustained 20%+ growth rather than simply protecting the installed base. The bull case is that Max increases wallet share and automation while preserving pricing power. The bear case is that growth decelerates faster than margins expand, making a premium multiple hard to defend. Watch Q3 platform growth, net retention, Max monetization and sales execution through the CRO handoff.
Sources: ServiceTitan results | CRO transition
3. Braze shows healthy enterprise demand and cash flow, yet after-hours weakness says AI monetization must also protect gross margin
Braze reported fiscal Q2 2027 revenue of $227.2m, up 26.2% year over year and above the roughly $220m consensus cited by market data providers. Remaining performance obligations were $1.09bn, dollar-based net retention improved to 110%, and customers with more than $500,000 of ARR rose to 361 from 282. Non-GAAP operating income increased to $22m from $6m and free cash flow reached $21.7m from $3.5m. Full-year revenue guidance increased to $910–913m, above prior guidance, but Q3 adjusted EPS guidance of $0.13–0.14 was below the roughly $0.16 consensus and non-GAAP gross margin slipped to 68.6% from 69.3%. Shares were down about 10% in after-hours trading shortly after the release.
The bull case is stronger than the stock reaction implies: retention is improving, larger customers are expanding and the company is converting growth into cash. The bear case is that AI-driven engagement can increase model, cloud and message-processing costs at the same time customers demand measurable ROI, limiting gross-margin expansion. This is an important read-through for customer-engagement and marketing software: AI can increase campaign volume while simultaneously reducing the value of human seats and increasing variable infrastructure expense. Salesforce, Adobe and Twilio face adjacent versions of the same debate. Investors should watch paid adoption of BrazeAI, gross margin, enterprise retention and whether the raised full-year revenue outlook survives the next two quarters without additional cost pressure.
Sources: Braze Q2 results | Braze investor relations | After-hours reaction
4. The software selloff after GPT-6 Astra is a new market reaction, not a new model story — and it raises the hurdle for application earnings
US application-software shares sold off again on September 8, with ServiceNow down roughly 5% and Salesforce and Intuit around 4% lower as investors revisited disruption fears following OpenAI’s GPT-6 Astra launch. The broader software and services group fell about 1.4%, while semiconductor names were comparatively resilient. Astra itself was covered previously; what changed yesterday was the market’s willingness to re-price software before there is evidence of near-term revenue displacement.
This matters because it creates an asymmetric earnings setup. Application vendors that merely show stable growth can still de-rate if investors believe frontier agents compress seats, implementation labor or workflow differentiation. The bull case is that systems of record, proprietary enterprise context and governed transaction layers become more valuable as agents act across them. The bear case is that many user-facing workflows migrate into the model interface, leaving software vendors with less pricing power even if their underlying databases remain necessary. The market may be overreacting to immediate revenue risk but underreacting to terminal-value uncertainty. ServiceNow, Salesforce, Intuit, Adobe, Workday and Atlassian are the most obvious public exposures; Datadog, Snowflake, MongoDB and cybersecurity platforms remain better positioned where AI increases machine activity rather than replacing human seats. The next catalysts are Adobe and Oracle on September 10 and subsequent evidence on seat growth, consumption and AI-specific monetization.
Source: Reuters market report
5. Boston Scientific converts cyber risk into an earnings warning — evidence that security failures can hit manufacturing, shipments and guidance directly
Boston Scientific disclosed that unauthorized network activity identified on August 25 caused a global outage affecting operating systems and business applications, impairing manufacturing and its ability to process and ship orders. In an SEC filing on September 8, the company said major distribution and manufacturing activities had largely resumed but that the disruption makes it unlikely to achieve its prior Q3 and full-year 2026 sales and adjusted-profit outlook. Boston Scientific had previously guided to 2026 adjusted EPS of $3.28–3.32 and sales growth of 5.5–6.5%; management plans to quantify the impact when it reports on October 28.
For technology investors, this is a cleaner cybersecurity spending signal than another threat-intelligence report because the incident has an explicit P&L mechanism: lost production, delayed orders, operational remediation and potentially working-capital disruption. The bull case for cybersecurity is that boards and regulators increasingly treat identity, endpoint, network and recovery controls as business-continuity infrastructure rather than discretionary IT spend. The bear case is vendor-specific attribution: Boston Scientific has not publicly identified the initial attack vector, so it is wrong to infer which security category failed. Palo Alto Networks, CrowdStrike, CyberArk, Zscaler, Rubrik and OT-security vendors nevertheless benefit from a broader resilience budget. The next hard datapoint is Boston Scientific’s October 28 estimate of lost sales, remediation costs and recovery timing.
Sources: Boston Scientific SEC filing | Reuters
6. ASML is solving a real High-NA bottleneck for large AI dies — strategically important, but the revenue payoff sits mostly beyond today’s forecast window
ASML said it will work with major chipmakers to adapt High-NA EUV for much larger data-center chips, addressing a technical limitation created by the smaller exposure field of the current High-NA platform. Existing EUV systems can pattern chips around 800 square millimeters, while large AI accelerators push toward the practical reticle limit. ASML’s roadmap uses larger masks and targets roughly 40% higher system productivity; Samsung and SK Hynix aim to use High-NA in DRAM by 2028, TSMC expects broader adoption around 2030, and ASML is targeting a larger-mask pilot line around 2031 with high-volume rollout around 2033. Separately, ASML broke ground on its BIC North manufacturing expansion in the Netherlands, with the first phase due in 2029; Reuters says nearly all current EUV capacity is booked through 2027.
The bull case is that AI chip size and pattern complexity make advanced lithography more rather than less valuable, extending ASML’s scarcity moat even as custom accelerators proliferate. The bear case is timing: the larger-mask roadmap is too far out to justify near-term estimate upgrades, and chipmakers will continue using multi-patterning, chiplets and advanced packaging to manage economics. The wider read-through is favorable for TSMC, Intel, Samsung Electronics and SK Hynix because High-NA can support denser logic and memory, while Nvidia and custom-chip designers gain a path to larger monolithic dies when economics warrant it. Investors should watch High-NA throughput, customer acceptance, yields and actual tool orders rather than roadmap milestones alone.
Sources: Reuters — High-NA roadmap | ASML BIC North | Reuters — manufacturing expansion
7. US accusations of Chinese model distillation raise the probability that AI controls move from chips to model outputs and APIs
US officials on September 8 accused six Chinese AI companies, including DeepSeek, Moonshot AI and Alibaba, of using “industrial-scale” distillation to copy capabilities from US models supplied by Anthropic, OpenAI, Google and xAI. These are government allegations, not adjudicated findings, and the Chinese embassy had not responded when Reuters published. The timing matters: the accusations arrive ahead of expected US-China leadership talks later this month and as Washington increasingly treats model capability itself, not only advanced chips, as a national-security asset.
The investor question is whether policy broadens from semiconductor export controls into restrictions on API access, model-output harvesting, cloud access or the distribution of high-capability weights. That would raise the cost of capability catch-up for Chinese model developers, but it could also accelerate domestic self-reliance and make globally open model ecosystems harder to sustain. OpenAI, Anthropic, Alphabet and xAI gain some protection against uncompensated distillation but face compliance and market-access costs; DeepSeek, Moonshot AI and Alibaba face the clearest downside if access becomes more restricted. Nvidia remains indirectly exposed because every escalation in model policy can spill into chip and cloud export rules. The next catalyst is any concrete US measure following the allegations, not the rhetoric itself.
Source: Reuters
8. OpenAI is moving enterprise AI toward vertical workflows and outcome-based pricing — potentially a bigger software threat than another model benchmark
OpenAI Chief Financial Officer Sarah Friar said at Goldman Sachs’ technology conference that the company is pushing deeper into specialized enterprise use cases including chip design, life sciences and financial services, while experimenting with pricing tied to business outcomes rather than raw usage. Friar said enterprise revenue grew faster than overall annualized revenue recently and that OpenAI used its own models to help develop its Jalapeno chip through tape-out in nine months. These are company statements rather than independently audited segment disclosures, but the direction is important: OpenAI wants to move up the stack from selling intelligence to capturing a share of the economic value created by completed workflows.
That is potentially more disruptive to enterprise software and services than lower token prices alone. Outcome pricing competes directly with consulting, outsourcing and workflow software where customers care about resolved tickets, completed designs or processed transactions rather than seats. The bull case for incumbents is that domain data, compliance, permissions and transaction systems remain necessary, giving companies such as ServiceNow, Salesforce, Synopsys and Cadence natural control points around the model. The bear case is that OpenAI becomes the orchestration layer and pushes application vendors toward lower-value data or execution roles. The next evidence investors need is disclosed enterprise revenue mix, repeatable vertical products and actual customer economics rather than anecdotal productivity.
Source: Reuters
9. China’s provisional duties on Japanese dichlorosilane show semiconductor materials are becoming another geopolitical lever
China imposed provisional anti-dumping measures on Japanese dichlorosilane from September 8, requiring importers to post cash deposits of 80.8% to 99.2% depending on the supplier. Dichlorosilane is a high-purity input used in semiconductor thin-film deposition for logic, memory and analog chips. China’s Ministry of Commerce says its preliminary investigation found dumping and material injury to domestic producers; Japan has protested the move. The measure remains provisional pending the final investigation, so it should not yet be modeled as a permanent cost increase.
The immediate global supply-chain impact is likely smaller than the headline tariff rate suggests because China can source from domestic and other Asian suppliers, but the strategic message is larger. Semiconductor trade friction is spreading from lithography tools and advanced GPUs into specialty chemicals, where qualification cycles are long and purity requirements are extreme. Shin-Etsu Chemical and Denal Silane face the direct commercial risk, while Chinese foundries and memory producers may face temporary qualification costs. Korean and domestic Chinese chemical suppliers could gain share. The bull case is accelerated localization and more resilient sourcing; the bear case is higher fab input costs and duplicated supply chains. Watch China’s final ruling, customer qualification of substitutes and whether other specialty materials become targets.
Sources: China State Council Information Office | Associated Press
10. AI infrastructure financing is tightening even as issuance explodes — balance-sheet quality is becoming a competitive advantage
Reuters Breakingviews, citing Goldman Sachs, estimates AI-linked debt issuance reached nearly $500bn by August 2026, roughly one-fifth of high-grade US debt issuance for the year. The new information is not that AI infrastructure requires capital; it is that lenders are becoming more selective as grid delays, permitting and supply-chain bottlenecks push revenue recognition further out. Project financiers increasingly want signed leases, permits and stronger collateral or guarantees before releasing capital, and yields on some collateralized transactions are rising. Some projects also rely on guarantees or economic support from ecosystem participants such as Meta or Nvidia.
This does not imply an AI demand collapse. It implies a transfer of economics toward companies with cheap capital and fully contracted power. Microsoft, Amazon, Alphabet and Meta have a structural advantage because they can fund projects internally or borrow at investment-grade rates. CoreWeave, Nscale, Lambda, Crusoe and private data-center developers face greater sensitivity to credit spreads, hardware depreciation and customer concentration. Vertiv, Eaton and server suppliers can see order timing move if project financing slips even when long-term demand remains intact. The bull case is that tighter underwriting weeds out speculative capacity and improves project quality; the bear case is that the marginal data-center build becomes uneconomic before utilization catches up. Watch credit spreads, covenant structures, grid-interconnection milestones and the share of projects backed by binding customer leases.
Source: Reuters Breakingviews
What to Watch
SailPoint reports fiscal Q2 2027 before the US market opens today, with the conference call at 8:30 a.m. ET; ARR, SaaS ARR, non-human identity adoption and the integration of Entro are the key cybersecurity datapoints.
Apple’s September event begins at 10 a.m. PT today. The investor read-through is less about launch-day unit forecasts than pricing, memory content, on-device AI, services positioning and whether higher component costs can be passed through without hurting upgrade demand.
Oracle reports fiscal Q1 2027 after the close on September 10. OCI growth, RPO conversion, data-center capex, financing intensity and free cash flow are the most important cross-stack tests for the AI infrastructure cycle.
Adobe reports fiscal Q3 on September 10. Investors need evidence that generative AI is improving core recurring revenue and monetization rather than only defending Creative Cloud engagement, with the additional backdrop of the Adobe Commerce security issue.
US CPI later this week matters unusually much for technology because the market is already differentiating between near-term earnings scarcity in semiconductors and long-duration software cash flows; a hotter print would raise the valuation hurdle for both, but especially for software and leveraged AI infrastructure.
Bottom line
The best way to read this morning is that AI demand remains strong, but the market is repricing who gets paid and who has to finance the build. Qualcomm–Amazon broadens the custom-silicon opportunity and strengthens the case that inference will be heterogeneous, while ASML’s High-NA work reinforces the long-duration scarcity of leading-edge manufacturing equipment. At the same time, nearly $500bn of AI-linked debt issuance and tighter project-finance standards show that even genuine demand can produce weak equity outcomes if capital intensity and funding costs absorb too much of the value.
Software is entering the same phase of differentiation. ServiceTitan and Braze both produced credible growth, retention and cash-flow metrics, yet the stocks were punished because forward expectations and unit economics matter more than a generic AI narrative. The most durable software exposures remain systems with proprietary operational data, machine-data infrastructure and cybersecurity enforcement; thin workflow layers and labor-heavy services face a higher terminal-value risk as models move toward outcome-based pricing. The central investor question is no longer whether AI grows. It is whether each layer can preserve pricing power and free-cash-flow conversion as intelligence gets cheaper and infrastructure gets more expensive.