1. Yesterday’s software tape produced perhaps the clearest evidence yet that the 2026 de-rating is becoming a stock-selection market rather than a blanket “AI kills SaaS” trade: Atlassian rose c.35% after hours while Datadog and HubSpot fell c.20%.
Atlassian delivered Q4 revenue of $1.77bn versus $1.66bn expected and adjusted EPS of $1.87 versus $1.50, while guiding the September quarter to $1.71–1.72bn, comfortably above the $1.66bn consensus. Conversely, Datadog fell more than 16% despite c.36% revenue growth, while HubSpot dropped c.22% after cutting FY26 revenue guidance by roughly $22m. What changed is the market’s willingness to discriminate between AI adoption and AI monetisation: Atlassian’s cloud migration, enterprise penetration and Rovo-driven automation are showing up alongside tangible revenue upside, whereas Datadog and particularly HubSpot are confronting uncertainty around consumption optimisation, agent pricing and the migration away from traditional seat economics. Bulls will argue the sell-off in DDOG/HUBS is another expectations reset rather than evidence of structural impairment; bears will argue that agents undermine the per-user pricing architecture underlying much of SaaS even when customer usage increases. The second-order implication is important for CRM, NOW, WDAY, ADBE, SNOW and MDB: simply launching agents will no longer protect multiples—the market increasingly wants evidence that AI raises revenue per customer faster than it reduces seats or unit pricing.
2. Cloudflare’s print may be the most strategically important software result of the week because it shows agentic AI creating an incremental traffic and security revenue pool rather than cannibalising the incumbent product.
Q2 revenue reached $696.1m versus $665.5m expected; Cloudflare guided Q3 to $736–737m versus $722.1m consensus and raised FY26 revenue guidance to $2.86–2.87bn from $2.805–2.813bn, sending the shares c.18% higher after hours. Management specifically attributed demand to businesses using Cloudflare’s network to route and secure rapidly expanding AI-agent traffic. This is the critical distinction from conventional application SaaS: an agent may eliminate human clicks, but it can generate more API requests, machine identities, inference calls and east-west traffic, all of which require connectivity, policy enforcement and protection. Bulls will argue Cloudflare is evolving from CDN/security vendor into an internet-scale control plane for agents and inference, expanding its opportunity across Workers, Zero Trust and security; bears will argue hyperscalers can internalise much of this functionality and that premium valuation already discounts substantial AI-driven acceleration. The read-across is particularly constructive for PANW, ZS, CRWD, CYBR, OKTA and DDOG: agentic AI increasingly looks deflationary for seats but inflationary for network traffic, security telemetry, observability and machine identity.
3. AMD’s acquisition of Taalas shows that the Nvidia challenge is shifting away from building a CUDA-equivalent GPU alone towards attacking the economics of inference with specialised silicon.
AMD agreed yesterday to acquire Taalas, whose architecture is designed specifically to reduce compute and memory bottlenecks in inference, following earlier acquisitions of MK1, MEXT and FastFlowLM. Taalas had raised c.$219m before the deal; AMD intends to integrate the technology into the Instinct roadmap and broader system-level AI platform. What changed is strategic emphasis. Training remains dominated by Nvidia’s general-purpose accelerator ecosystem, but the much larger long-term compute opportunity may ultimately be serving trillions of recurring inference requests, where latency, memory bandwidth, power and cost per token matter more than absolute training flexibility. Bulls on AMD will argue that heterogenous inference gives the company a much larger attack surface than simply replacing Nvidia GPUs one-for-one and allows AMD to combine EPYC, Instinct and specialised accelerators. Bears will argue that Nvidia recognises exactly the same shift—its Groq-related technology strategy is designed around inference—and continues to control networking, software and system architecture. Second-order beneficiaries include TSMC, HBM, AVGO, MRVL and advanced packaging, but the larger implication is potentially negative for accelerator ASPs: if inference becomes workload-specific, the market may fragment across GPU, ASIC and specialised architectures rather than remaining a single premium general-purpose compute pool.
4. Tesla and SpaceX’s planned
$16.8bn Terafab is a meaningful escalation in vertical integration and signals that large AI users increasingly view semiconductor capacity itself as strategic infrastructure. The companies plan an initial $16.8bn investment in a Texas semiconductor complex spanning manufacturing through testing, with potential investment ultimately reaching roughly $119bn; the stated requirement is to support more than 1 terawatt of computing demand across Optimus, Cybercab and SpaceX’s prospective space-based data centres. This matters less because Tesla suddenly becomes a credible TSMC competitor—the technical and execution barriers remain immense—and more because AI customers are responding to scarce leading-edge compute by owning progressively more of the supply chain. Bulls will see another confirmation that semiconductor demand extends well beyond Microsoft/Amazon/Google/Meta and that sovereign and vertically integrated buyers can prolong the infrastructure cycle. Bears will see the opposite: every major customer is now simultaneously signing long-term supply agreements, designing chips and financing new fabs, which increases the probability of substantial late-decade overcapacity. The immediate read-across is mixed for INTC—which is reportedly involved in supporting the manufacturing effort—and longer-term relevant for TSMC, Samsung, NVDA, AMD and AVGO. Washington’s simultaneous decision to impose a 15% tariff and price floors on Chinese-linked polysilicon reinforces the policy direction: domestic AI compute capacity is increasingly a national-security objective rather than purely an ROIC decision.
5. The most important semiconductor signal from yesterday was not weak AI demand but collapsing tolerance for anything short of persistent scarcity: Western Digital fell c.19%, Sandisk c.13%, with weakness spreading into Micron, SK Hynix and the broader chip complex despite very strong operating results.
Sandisk’s data-centre revenue increased roughly 400% yoy, while both storage vendors beat consensus and continued to point to exceptional AI infrastructure demand; nevertheless, investors focused on indications that pricing growth could eventually normalise after extraordinary share-price appreciation—Sandisk had risen more than fivefold and Western Digital roughly threefold. This is increasingly the pattern across AMD, storage and memory: fundamentals remain strong, but stocks are starting to discount the derivative of growth rather than growth itself. Bulls will argue supply remains contractually tight, AI storage requirements are exploding and the current sell-off is simply multiple digestion. Bears will argue the equity market is beginning to look through 2026–27 scarcity towards an eventual synchronised supply response across NAND, DRAM/HBM, foundry, packaging and data centres. The distinction matters for positioning: NVDA and AVGO arguably deserve better durability premiums because ecosystem and custom-silicon value is less directly commodity-price dependent; MU, SK Hynix, Samsung, WDC and SNDK remain more sensitive to the duration of scarcity. For the equipment names—LRCX, AMAT, KLAC, ASML—the paradox continues: the more profitable today’s shortages become, the more aggressively customers invest to eliminate them.
Bottom line
yesterday strengthened rather than weakened the AI investment thesis, but it sharpened where value is accruing. The market rewarded control points—Atlassian workflow, Cloudflare network/security—and punished software where monetisation visibility is weaker, while semiconductor investors increasingly distinguish long-duration ecosystem rents from cyclical scarcity. The most interesting second-order trade remains the same: agents threaten seat-based SaaS economics while simultaneously increasing spend on networking, observability, identity and cybersecurity.
