The Semiconductor Earnings Boom Is Just Getting Started | Ben Pouladian on why AI is Real, Nvidia is Mispriced, and Capacitors Are Overrated

Ben Pouladian argues Nvidia and compute infrastructure differ fundamentally from the Cisco and fiber buildout of the late 1990s because compute generates intelligence from nothing, whereas fiber only moves data between two points. He counters the commodity-AI objection by distinguishing frontier model tokens, which he compares to jet fuel, from free smaller model tokens, which he compares to water, with Meta's 30 billion parameter Llama representing effectively commoditized low-value output. Nvidia revenues grew from approximately 27 billion dollars in 2023 to 250 billion dollars over the last twelve months, and Pouladian estimates operating profit over the next twelve months conservatively at 200 billion dollars pre-tax. He emphasizes that semiconductor stock gains over this period were driven entirely by earnings growth rather than multiple expansion, with Nvidia currently trading at a forward multiple below its own five to ten year historical average, which he interprets as the market pricing in declining pricing power rather than a bubble.

The bearish case Pouladian acknowledges is that Google, Amazon, Meta, and ByteDance are all developing custom accelerators to avoid the Nvidia tax, and the market expects margins to decline from extremely high to merely very high. He counters that Google, despite full internal access to its own TPU silicon, still buys billions of dollars per year of Nvidia GPUs and sells Nvidia-based compute on Google Cloud because demand exceeds supply. He argues Nvidia is more than half software and coordinates hardware and software co-design with partners including Anthropic and OpenAI, and that any hardware company without a software layer delivering ongoing customer benefit will commoditize over time.

The current infrastructure constraint is not chip fabrication, which Pouladian says is fully automated at Taiwan Semiconductor, nor a GPU shortage, but a shortage of powered land and skilled tradesmen to build data centers fast enough. The lag in buildout is measured in years not quarters. Bloom Energy is identified as a notable power play because it deploys modular fuel cells at a data center site in under 80 days delivering 800-volt DC power without transformers, and data center developers borrowing at seven to eight percent are incentivized to use it to start charging customers rather than waiting for utility connections. GE Vernova's turbine backlog extends to 2028 or 2029, though Pouladian views turbines as a temporary solution, and predicts future data center power will come from a combination of natural gas, solar, batteries, and conventional nuclear, with small modular reactors less likely than traditional nuclear.

ASML holds a literal monopoly in EUV lithography and Lam Research operates in an oligopolistic market, making them less risky than leveraged neocloud borrowers, several of which Pouladian says are highly leveraged and some of which will fail or be absorbed. Approximately 30 percent of Lam Research revenue is recurring from spare parts and software, and he predicts that stream will continue into the 2030s even as headline growth normalizes. He flags a valuation tension, arguing Nvidia cannot trade at a low multiple while semi cap names simultaneously trade at high multiples, implying mean reversion is required somewhere.

Pouladian estimates Anthropic inference gross margins in the 70 percent range or higher, comparable to Salesforce SaaS margins of 72 to 77 percent, and argues inference profitability is real. Anthropic reached one billion dollars in annual revenue run rate by February. He describes an emerging enterprise router approach where hard tasks use expensive frontier models and simpler tasks use local open source models, citing Nvidia itself using Anthropic Opus to orchestrate and its own open source model Nemotron for routine tasks. He argues Nvidia developed Nemotron partly to prevent customers like Anthropic from vertically integrating into chip design, and notes that because Nemotron is optimized on CUDA, global open source community contributions improve Nvidia's ecosystem at no cost to Nvidia.

Pouladian is skeptical of capacitors and resistors as an investment theme, predicting chasing them will probably end in tears, and notes Vichay is already down 35 percent from its late June peak. He draws a sharp contrast with HBM memory, where only three players exist and the manufacturing process requires nanometer-precise stacking of multiple wafer layers through silicon vias, a technical barrier that prevented manufacturers from advancing beyond 14 layers when attempting to move to 16. Commodity capacitors face no equivalent barrier and are vulnerable to Chinese government-funded competitors flooding the market.

Pouladian expects Micron, SK Hynix, and Samsung to print hundreds of billions of dollars in profit over roughly the next year, but emphasizes the most money is made by buying before the profitable period begins rather than after pricing power starts declining. He anticipates memory prices will begin declining around early 2028, with market reaction likely becoming visible around mid-2027, driven by fabs coming fully online running three shifts and flooding supply. He identifies the next major AI inflection point as the intersection of artificial intelligence with material science and biotech, arguing AI could compress ten years of laboratory research into one year for drug discovery and genomics, and frames AI leadership as a national security imperative analogous to the space race, with China replacing Russia as the primary competitor.