Notes from Silicon Valley

I was excited to be back in Silicon Valley again this year. Danske Bank had set up an ambitious, high-quality meeting agenda for a small group of Nordic investors. During the week I sat down with Nvidia, Applied Materials, Arista, Cisco, Cadence, Autodesk, DocuSign, Visa, Sandisk, Western Digital, Bloom Energy, Rigetti Computing and one interesting newcomer, Majestic Labs.

If last year’s trip was about confirming that AI demand was real, this year was about something more specific. Almost every conversation, regardless of subsector, circled back to the same handful of physical, financial and structural constraints. And software is not dead!

Nvidia’s own framing was blunt: the limiting factor is no longer GPU supply but ”land, power and shell”. Suitable sites, electricity and completed data-center buildings. Demand still exceeds what Nvidia can ship (guiding for 70% revenue growth in FY2028 that could have been 100% without constraints), but CFO Kress admitted that memory, substrates and advanced packaging are now just as big a constraint as compute itself.

Applied Materials has raised its 2026 equipment growth outlook three times this year, driven by leading-edge foundry, DRAM and advanced packaging. The company has doubled its manufacturing footprint and is now selling on value rather than price for the first time in years. Bloom Energy is benefiting from the energy bottleneck: its solid-oxide fuel cells aren’t the cheapest power on a pure €/kWh basis, but they deploy in months rather than years and one project went from signed to online in 55 days! Speed to power, not cost of power, is becoming the currency.

Memory has ”quietly” become the new battleground. Sandisk described NAND’s shift from a brutally cyclical, customer-controlled commodity into a strategically contracted business, with visibility rising from roughly three months to as much as four years in two quarters, as hyperscalers post cash and credit guarantees to lock in supply. Western Digital told a similar story on hard drives, with agreements now stretching to 2029-2031 and gross margins above 54%. Both are turning a volatile commodity into something stable and predictable.

Then there was Majestic Labs, a pre-revenue startup founded by ex-Google/Meta chip engineers betting that HBM-centric GPU architecture is already hitting a ”memory wall” for very large inference workloads. ML believes that separating compute from a large pool of cheap memory (LPDDR) could handle the same workloads at a fraction of the power and price. Nothing is proven yet (first customer silicon isn’t expected until 2027), but it’s a useful reminder that the AI accelerator market may fragment by workload rather than consolidate around one architecture.

Networking is turning into a security-driven refresh cycle. Arista and Cisco offered two sides of the same coin, disagreeing almost everywhere else (merchant silicon vs. proprietary Silicon One; open partnerships vs. owning the security stack), but both pointed to the same trigger. Frontier AI models (especially Mythos) can find vulnerabilities in old, unsupported network infrastructure. This has turned a slow, discretionary refresh into an urgent security mandate. AI isn’t just a demand driver for infrastructure anymore, it’s starting to actively expose the infrastructure that hasn’t kept up. Visa echoed this from the other side. Through the same Glasswing consortium giving companies early access to Mythos, it stress-tested its own systems, found its defenses held, and open-sourced the resulting testing harness as a new consulting offering.

On the software side, the shared message was that owning the data and the authority to act on it matters more than owning a model. DocuSign, Autodesk, Cadence and Visa each made a version of this argument. General-purpose AI is commoditizing, but none of it substitutes for proprietary project data, legal authority to execute, decades of deterministic engineering logic, or a trusted transaction network. As Docusigns CEO put it, frontier models may commoditize generic intelligence and ease technical integration, but they don’t gain access to confidential agreements, internal approval structures or regulatory obligations. Autodesk is betting its 40-year engineering history is worth more than a generative image model, and Cadence argues that agentic chip-design tools actually increase demand for its software, since agents still have to call on Cadence’s core engineering tools to check and verify their results. Hinge Health, meanwhile, emphasized the value of its proprietary patient and treatment data.

Another observation is the financial engineering layer that is emerging across the industry. Nvidia now backstops up to 25% of certain neocloud financing arrangements, helping customers secure lower rates. Bloom relies on partners like Brookfield to bridge the gap between how customers want to pay (per kWh) and how Bloom wants to be paid (equipment revenue, upfront). Sandisk and Western Digital’s long-term contracts increasingly resemble structured credit instruments. What used to be a pure hardware and demand story now also bake in balance-sheet risk, which is worth remembering.

Geopolitics and sovereignty came up in nearly every room, which wasn’t a surprise. Nvidia and Arista discussed sovereign AI opportunities tied to neocloud demand and vendor-access considerations. Export controls and China exposure surfaced at Applied Materials, Cadence and Sandisk. But sovereignty concerns also came up at Visa and DocuSign, both facing early European competition from domestic-preference procurement. Multi-sourcing, local compliance and ”what if a government cuts our access” scenario-planning are now standard agenda items well outside the semiconductor sector. Visa fielded direct questions about whether a U.S. administration could threaten access to its infrastructure? IR called this an unavoidable risk of being U.S.-headquartered, adding that it works with governments while operating locally wherever legally permitted.

Quantum computing, as topical as it is, still feels years away, and increasingly state-backed. Rigetti was refreshingly honest that quantum computing remains a milestone-driven engineering story with qubit count, fidelity, and coherence time, but it’s not yet a market. The US Department of Commerce recently took a small equity stake tied to roadmap milestones, proof that quantum’s strategic importance in Washington now rivals its technological readiness.

So in summary, the debate this year wasn’t “AI hype vs. reality”, that seems settled. It was about the next constraints to solve: power, memory, security, and the capex and financing holding it all together. Real, deployable solutions are however emerging for each constrain with faster power, smarter memory architectures and more urgent security refreshes. The chip isn’t the sole story anymore. Everything around it is.

I wrote about self-driving cars in my previous Silicon Valley report. The fleet has roughly tripled across the Bay Area since then, and the driving has changed too. A year ago I could spot a Waymo by how it drove, but now it’s often hard to tell until you see the empty driver’s seat and the sensors. A small but telling example of AI getting better through real-world experience.

Portfolio Manager
Janna Haahtela