Foundation Model Valuations Near $1 Trillion But History Says Infrastructure Builders Rarely Win

InnovationVenture CapitalFoundation Model Valuations Near $1 Trillion But History Says Infrastructure Builders Rarely WinByJosipa Majic Predin, Contributor. Forbes contributors publish independent expert analyses and insights. I’m a founder, writer and lecturer focusing on VC funds.Follow AuthorJun 03, 2026, 09:56am EDT--:-- / --:--This voice experience is generated by AI. Learn more.This voice experience is generated by AI. Learn more.The Microsoft and OpenAI logos are seen on screens in this illustration photo in Warsaw, Poland on 21 November, 2023. Former head of OpenAI Sam Altman has said he is still willing to lead the company after his ousting if two board members resign according to The Verge. (Photo by Jaap Arriens/NurPhoto via Getty Images)NurPhoto via Getty ImagesAnthropic closed a $65 billion Series H in late May at a $965 billion valuation, leapfrogging OpenAI's $852 billion mark set just two months earlier. Together the two largest foundation labs have absorbed 14% of all global venture capital in 2025. The last time a single technology category commanded this share of global private capital and this level of infrastructure-layer excitement was the late 1990s, when Nortel Networks carried 75% of North America's internet backbone traffic and peaked at a $366 billion market cap. In 2009, it filed for bankruptcy. The structural parallel may be warning about where value accumulates in infrastructure cycles. Foundation model funding doubled in Q1 2026 versus all of 2025, with capital increasingly concentrated in a handful of labs. Anthropic’s revenue trajectory is genuinely unprecedented: the company crossed $30 billion annualized run rate by early April 2026, having generated roughly $9 billion at the end of 2025. Salesforce took two decades to reach that figure. But revenue growth at the infrastructure layer does not determine who captures the value the infrastructure enables, and investors are beginning to price in a different question. Om Malik, the iconic San Francisco-based writer and investor, argued this week that the right historical analogy for AI’s current moment is not the iPhone plateau or the PC clock-speed wars. It is Dense Wavelength Division Multiplexing; the optical networking technology that made the internet functionally unlimited in capacity while simultaneously disappearing from public conversation. DWDM, built primarily by Nortel and Lucent in the mid-1990s, enabled YouTube, Netflix, and Zoom but it did not define who profited from them. The mechanism Malik identifies is commoditization-through-capability. Per-token inference costs have fallen 280-fold between 2022 and 2024, according to Stanford’s 2025 AI Index Report. Epoch AI’s analysis puts price-per-token declines at between 9x and 900x per year depending on the performance benchmark. When the cost of capability drops that fast, the capability stops being the competitive moat: the application becomes it. This is precisely what happened to optical networking. Ciena, one of DWDM’s pioneers, survived the dot-com crash while Nortel and Lucent collapsed; not because DWDM failed, but because the companies building on top of it (Google, Amazon, Netflix) captured the value the infrastructure made possible. The infrastructure builder's dilemma is structural: the more successful you are at commoditizing capability, the more you compress your own margins. MORE FOR YOUOpen-weight models are already accelerating this dynamic. The gap between frontier closed models and open-weight alternatives has narrowed materially since 2023, compressing the pricing power that closed labs once held. Enterprise AI budgets grew from $1.2 million annually in 2024 to $7 million in 2026, but that spending growth increasingly flows to inference and application infrastructure rather than to foundation model API fees. The unit economics that justified frontier-lab valuations depend on maintaining a performance gap that open-weight models are systematically closing. Malik’s DWDM analogy carries one further implication that his piece does not fully surface: the companies that built the optical backbone were not stupid or badly run. Nortel’s optical networking revenues soared 133% to $9.2 billion in 2000. The problem was that the infrastructure they built made them replaceable. The question for AI investors in 2026 is whether foundation labs have built enough application-layer lock-in: through developer ecosystems, proprietary data flywheels, or enterprise relationships, to avoid the same structural fate. The signal Malik points to watch the disappearance of the benchmark from conversation. When enterprise buyers stop asking which model scores highest on reasoning tests and start asking why their software feels smarter, the infrastructure transition will be complete. That transition is already underway, OpenAI filed its IPO prospectus confidentially in May 2026, with Anthropic likely to follow. Both are racing to the public markets before the window closes, before investors internalize that the value the optical fiber enables is not the same as the value of the optical fiber itself. Investors and founders building at the application layer should read this moment carefully. The companies that will define the AI era are probably not the ones currently consuming 14% of global venture capital. They are more likely the ones quietly building on top of it: the Googles and Amazons to this cycle's Nortel. The optical fiber is already in the ground. Editorial StandardsReprints & Permissions