Highlighted Companies
KEY COMPANIESAt Dreamforce (September 15–17), Salesforce unveiled AIforce and Koa — its first CRM reasoning model, built on NVIDIA Nemotron — put Claudeforce (Anthropic) into open beta, and wired Agentforce to Google’s Gemini Enterprise over MCP and to AWS. Dario Amodei and Jensen Huang shared Benioff’s stage, and Salesforce disclosed Agentforce ARR up more than 240% toward a ~$1.5B FY27 target. The read-through for the accelerator names: the enterprise-app layer is now a demand pull-through for NVIDIA silicon and frontier models alike, not a competing layer.
Read More →A week after posting 62% cloud growth and a $664B RPO, Oracle fell roughly 3% on September 18 and sits down about 24% year to date. The concern isn’t backlog or demand; it’s the interest expense and heavy capex required to fund the buildout, a worry the Fed’s September 16 hike only sharpens. Oracle is the clearest example of the week’s dividing line: the AI infrastructure story is intact, but the cost of financing it is now being priced as a real risk on the most leveraged builders.
Read More →Microsoft completed the migration of Copilot into a single unified product, shipping Windows and Mac desktop apps by mid-September and advancing multi-step autonomous “AutoPilot” agents. Paired with ServiceNow surfacing its AI specialists in the new Microsoft Agent 365 Marketplace, Microsoft is quietly assembling the distribution layer for third-party agents — the same governed control-layer position Salesforce and ServiceNow are racing toward, but anchored to Windows and the Microsoft 365 install base.
Read More →AI & Semiconductors
AI & CHIPSDays after Nvidia moved Vera Rubin into mass production, CoreWeave announced a multi-rack Rubin NVL72 cluster running in its cloud — the first live deployment, built on TSMC’s 3nm process with HBM4 memory. The speed from mass production to a running cluster is the datapoint that matters: it validates that the ramp behind Nvidia’s $40B-per-gigawatt economics is real, and it puts CoreWeave ahead of the hyperscalers in showcasing next-generation silicon at rack scale.
Read More →HBM4 pricing is projected to climb from about $2 per gigabit to $4–5 or higher in the second half of 2026, and Nvidia alone has reserved roughly 800,000–850,000 wafers — more than 50% — of TSMC’s CoWoS advanced-packaging capacity for the year. With the supply-demand gap still near 20% even as CoWoS heads toward 130,000–140,000 wafers per month by year-end, memory and packaging — not the GPU die — are now the binding constraint on how fast AI compute can actually ship.
Read More →Cerebras demonstrated a combined pipeline in which AMD Helios GPU racks handle high-throughput context and prompt processing while Cerebras Wafer-Scale Engines take latency-sensitive token generation. It is the clearest sign yet that the AI hardware race has become a systems race: rather than one chip winning outright, inference workloads are being split across specialized silicon — a structure that plays to AMD’s full-stack push and chips away at Nvidia’s assumption that one platform serves the entire workload.
Read More →Markets & Tech Stocks
S&P 500 · NASDAQ · MARKETSThe FOMC voted 12–0 on September 16 to raise rates 25 basis points to 3.75–4%, the first increase since 2023, and 16 of 18 officials see at least one more hike this year. Markets now price roughly a 90% chance of a further hike by the December 9 meeting and about 55% for October 28. For an index carried by a handful of long-duration AI names, a rising discount rate is the single biggest threat to index-level returns regardless of how strong earnings look.
Read More →After an initial drop on the decision, stocks rebounded hard in the following session as softer oil, lower Treasury yields, and strong jobs data eased rate-hike fears — the Nasdaq jumped 1.7% to 26,418.30, the S&P 500 rose 1.1% to 7,637.76, and the Technology Select Sector SPDR advanced 2.2%. The pattern is now familiar: the AI-capex trade keeps absorbing macro shocks and bouncing, because the buildout momentum behind it has not broken.
Read More →Friday’s quarterly “triple witching” expiration brought selling and rising yields, leaving the week split: the rate-sensitive Dow fell 1.7%, the S&P was roughly flat at -0.1%, and the tech-heavy Nasdaq eked out +0.7%. The dispersion is the story — capital is concentrating in the AI-exposed, cash-generative names and leaving the rate-sensitive rest of the market behind, exactly the rotation a hiking Fed tends to force.
Read More →Supply Chain & Commodities
CHIPS · MATERIALS · FREIGHTTSMC’s CoWoS capacity is projected to reach 130,000–140,000 wafers per month by the end of 2026, yet demand still exceeds supply by roughly 20%, and HBM4 prices are set to double to $4–5 per gigabit in the second half. With Nvidia holding more than half of 2026 CoWoS allocation, the constraint on AI infrastructure has moved decisively from GPU dies to the memory and packaging around them — the layer least able to add capacity on the timeline the capex plans assume.
Read More →Digitimes’ weekly roundup flags HBM4 supply strain, fresh Intel price increases, and the next leg of the AI infrastructure buildout — a reminder that TSMC’s planned up-to-10% price increases from 2027 and tightening memory are shifting margin from fabless chip designers back to the foundries and memory makers. For every AI hardware vendor guiding multi-year revenue targets on stable gross margins, the supply chain is quietly repricing the inputs underneath them.
Read More →