Reconstructed using information available around the failed 07:52 UK scheduled run.
1. Kimi K3 has already complicated the “efficient models destroy compute” thesis: demand has overwhelmed Moonshot’s infrastructure.
Moonshot paused new Kimi subscriptions after usage exceeded its available compute capacity, only days after releasing the 2.8tn-parameter open-weight model. This is the most important incremental development because it cuts against the simplest DeepSeek-style bear case. Kimi K3 may commoditise the model layer and pressure the pricing power of OpenAI, Anthropic and other proprietary labs, but cheaper access appears to be stimulating enough usage to create an immediate infrastructure bottleneck. The investor debate should therefore distinguish compute per task, which is falling, from aggregate compute demand, which can still rise as lower prices unlock more users, agents and workloads. The likely winners are inference capacity, clouds, custom silicon, memory, networking and observability; the more vulnerable layer is proprietary-model economics rather than semiconductor demand in aggregate. Most exposed: NVDA, AMD, AVGO, TSMC, MSFT, GOOGL, AMZN, META, Alibaba, DDOG and private OpenAI/Anthropic.
2. Semiconductors have entered a bear market despite outstanding earnings, showing that the market is now discounting duration rather than present demand.
The Philadelphia Semiconductor Index ended Friday just over 20% below its late-June peak, while TSMC fell roughly 7% despite reporting a substantial earnings beat and raising investment plans. The semiconductor earnings season should still produce strong profit growth, but investors are asking whether hyperscaler capex, HBM pricing and leading-edge utilisation can remain exceptional beyond 2027. This is a major change in the debate: good results no longer automatically drive stocks because the market increasingly fears that extraordinary margins are inducing extraordinary capacity. Rising oil prices above $90/bbl and the US 30-year yield moving above 5% also worsen the valuation and data-centre-cost backdrop for long-duration AI assets. Most exposed: NVDA, MU, AVGO, AMD, TSMC, ASML, AMAT, LRCX and KLAC; relative beneficiaries from a rotation could include hyperscalers and software platforms monetising already-installed compute.
3. Alphabet is this week’s cleanest test of whether hyperscalers can convert unprecedented AI capex into acceptable revenue and returns.
Alphabet reports on 22 July, having previously guided to $175–185bn of 2026 capex, almost double its 2025 expenditure. The hurdle is no longer proving that Google needs more GPUs and data centres; investors need evidence that Google Cloud, Gemini usage, enterprise inference and advertising productivity are growing quickly enough to absorb higher depreciation, power and infrastructure costs. Strong cloud growth and improving AI monetisation would support GOOGL and reopen the broader AI supply chain; weak conversion would reinforce the argument that suppliers are capturing the economics while hyperscalers bear the capital burden. The second-order implication is that AI market leadership could rotate from semiconductor scarcity towards platforms with distribution, proprietary data and the ability to meter inference. Most exposed: GOOGL, NVDA, AVGO, ANET, VRT, MU, MSFT, AMZN and ORCL.
4. ServiceNow and SAP will determine whether software’s recent stabilisation is a genuine fundamental turn or merely relief from oversold valuations.
ServiceNow reports on 22 July and SAP on 23 July, directly after IBM warned that customers were redirecting spending towards servers, storage and expensive memory; IBM’s preliminary Q2 revenue of $17.2bn was below the $17.86bn consensus estimate. The core SaaS debate is therefore broader than seat cannibalisation: AI infrastructure may consume enterprise technology budgets before application vendors generate enough agent revenue to compensate. ServiceNow needs to show that AI agents increase workflow consumption and platform value, while SAP needs evidence that Joule and cloud migration deepen customer economics rather than simply raise delivery costs. Better-positioned software remains usage-based infrastructure, data and observability; traditional application vendors still need to prove that AI attach exceeds seat and budget pressure. Most exposed: NOW, SAP, CRM, WDAY, ADBE and TEAM, versus DDOG, SNOW, PLTR and infrastructure software.
5. Cybersecurity remains the strongest relative software narrative because AI is turning identity, access and verification into compulsory infrastructure.
Recent incidents at Abbott and Clover Health reinforced the continued rise in AI-assisted attacks and ransomware, while the White House is establishing a coordination group connecting frontier-model developers with essential-services providers to share vulnerabilities identified by advanced AI systems. The emerging “synthetic insider” threat—attackers using stolen identities, deepfakes and remote-worker impersonation—broadens the cyber opportunity beyond endpoint detection into identity verification, behavioural monitoring, privileged access, data-loss prevention and zero-trust enforcement. The investor debate is less about whether demand grows and more about who consolidates it: platforms controlling telemetry and enforcement points should capture the largest budgets, while point tools risk bundling pressure. Most exposed: PANW, CRWD, ZS, OKTA, CYBR and FTNT; second-order beneficiaries include TENB, QLYS, RBRK and CVLT.
