AI, Semiconductors, and Google: Rethinking Valuations

This article distills a recent LinkedIn Live event hosted by Marvin Labs. The discussion brought together Alex Hoffmann (Co‑Founder & CEO, Marvin Labs), Max Stamakun, CFA (Co‑founder & Portfolio Manager, Israilov Financial LLC), and moderator James Yerkess (Former Global Head of Transaction Banking & FX, HSBC Wealth Management).

The panel focused on three topics that matter for investors today: what the current AI wave means for semiconductor capex and depreciation, how Google’s AI pivot interacts with search economics, and why many research frameworks need a refresh. Watch the full conversation below.

Takeaway 1: AI is reshaping the semiconductor capex cycle—and the depreciation math

Panelists agreed the industry is in an intense investment phase. Hyperscalers are directing large capex toward GPUs, networking, data centers, and high‑bandwidth memory (HBM). The pace of product cycles is accelerating, highlighted by NVIDIA’s Blackwell platform and signals of a faster cadence for flagship accelerators NVIDIA Blackwell.

The capex surge is only half the story. The harder question is depreciation. Several large platforms extended useful lives for servers and networking equipment in 2022–2023, which helped reported earnings. Microsoft moved from 4 to 6 years effective FY23 Microsoft FY23 10‑K, Alphabet extended servers and networking lives in 2023 Alphabet 2023 10‑K, and Amazon lengthened server lives in 2023 Amazon 2023 10‑K. Those changes made sense for general purpose compute. They are much harder to justify for leading‑edge AI accelerators that face rapid obsolescence.

It is hard to square a five‑year depreciation schedule with chips that can be outclassed within 12–24 months in both capex and opex per unit of compute.

For research models this implies: re‑examine useful life assumptions for AI hardware, pressure‑test DSG&A and COGS sensitivity to accelerated refresh, and be explicit about the capex‑to‑opex trade‑offs from migration to newer nodes and platforms.

Where value may accrue in the stack:

• Cutting‑edge accelerators remain the economic center of gravity, but availability is tight and cycles are shortening.
• HBM suppliers are strategically positioned as memory bandwidth becomes the bottleneck. SK hynix and Micron have announced HBM3E supply for leading AI platforms SK hynix HBM3E, Micron HBM3E.
• Semiconductor manufacturing equipment retains structural importance. ASML remains the sole supplier of EUV lithography systems used for advanced nodes ASML Annual Report.
• Foundry concentration continues. TSMC’s US build‑out shows the complexity and timeframes involved. The company has publicly guided to phased ramps and revised schedules in Arizona amid workforce and permitting challenges TSMC Arizona updates.

Takeaway 2: Geopolitics is now a core semiconductor variable

Export controls remain a swing factor for AI hardware flows into China. The US Bureau of Industry and Security has tightened rules on advanced computing and semiconductor equipment since late 2023 and updated them again in 2024 BIS Oct 2023 rule, BIS 2024 update. Panelists expect continued policy risk around leading‑edge accelerators and re‑exports through third countries. The China domestic ecosystem has scale in mature nodes, but panelists see a longer path to parity at the top end.

For investors, the portfolio implication is clear: scenario analysis for export rules, foundry geopolitics, and friendshoring is now table stakes. Model chassis should include explicit policy shock scenarios and their flow‑through to capex, depreciation, and gross margin.

Takeaway 3: Google’s AI pivot—cannibalization risk vs monetization upside

Alphabet’s AI quality has improved and Gemini is competitive at a model level, but consumer adoption and product distribution are the near‑term hurdles. The key question is not model parity. It is how AI changes the economics of search and ads.

• Cannibalization: AI overviews can compress query volume and downstream clicks. Alphabet has not disclosed click‑through below AI answers, which leaves a real uncertainty for search unit economics.
• Monetization upside: AI can improve targeting, auction efficiency, and yield across Google’s inventory. Advertising still drives the majority of Alphabet revenue Alphabet 2023 10‑K. Small improvements in pricing and conversion can be material.
• New models: Consumer chat products monetize differently than search. A $20 per month subscription implies far higher ARPU than ad‑supported search for a subset of users OpenAI ChatGPT Plus pricing. Ads and commerce integrations for AI chat are still early.

Google has become a stock for nobody. If you believe AI will raise the value of ad inventory, there are cleaner ways to express that thesis than owning a business where search could be cannibalized.

Panelists contrasted Alphabet with Meta and Apple. Meta has a direct path to monetizing better targeting at unchanged surface area. Apple could use on‑device AI to raise perceived hardware value and average selling prices without matching hyperscaler capex. Both routes avoid the search cannibalization overhang.

Takeaway 4: Update research frameworks before the cycle updates you

Two areas need attention in standard models:

• Useful life and refresh cadence: Many large platforms extended equipment lives during the cloud era. Analysts should test shorter lives and accelerated refresh for AI accelerators and networking, then flow that through to D&A, capex intensity, and free cash flow. Cite company‑specific policies from primary sources and avoid applying cloud‑era assumptions to AI clusters Microsoft FY23 10‑K, Alphabet 2023 10‑K, Amazon 2023 10‑K.
• Policy sensitivity: Incorporate export control scenarios, local content rules, and friendshoring dynamics for foundry capacity and equipment availability BIS Oct 2023 rule, TSMC Arizona updates.

What to watch next

Closing thoughts

The panel’s message is straightforward. The AI build‑out is real, the depreciation math is not settled, and policy risk is now embedded in the semis value chain. For analysts, that means revisiting useful life assumptions, mapping policy scenarios into models, and leaning on tools that can parse primary sources at scale.

To hear the full debate and the nuances behind each viewpoint, watch the complete LinkedIn Live recording above.