AI demand still runs ahead of supply, but the compute bottleneck is turning into a political problem
Glen Kacher of Light Street sees AI as a 15-to-20-year rebuild of the computing stack, and only a third of the way through the first phase; the real risk is not over-investment in CapEx but power and local politics slowing construction.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Buy only the best company, never the second or third
Over 13 years at Integral Capital, Kacher made 46 private investments, and the lesson he drew is: do not buy the second- or third-best company in an industry, always buy the best one you can get. The reason is that in tech the leader takes two-thirds of the market, the runner-up gets 20% to 25% at most, and everyone else fights over scraps. Higher margins plus a dominant share compound the lead. But he also concedes that incumbents in tech get disrupted, and disruption usually comes not from big companies but from small ones — AI and the semiconductors behind AI are exactly the test case for that rule.
— Glen KacherNvidia's moat is underappreciated
Kacher says many people assume Nvidia will lose its roughly 85% share of the AI accelerator market, and the shift to inference does give competitors an opening. But he thinks the market currently underappreciates Nvidia's own capacity to innovate. He describes the evolution of inference chips as an alphabet progression from CPU to FPU to GPU to XPU, where X stands for the unknown: Google has TPU, there are options like Trainium, and Nvidia itself acquired the Grok inference line. He judges that inference will ultimately be a bigger market than training, so there will be many chip forms and many opportunities to innovate.
— Glen KacherFour semiconductor companies are 40% of the portfolio
TSMC, Nvidia, Broadcom and AMD make up roughly 40% of Light Street's public holdings, a highly concentrated portfolio. Kacher's logic: concentrate capital where innovation is densest, and right now the densest place is the core of AI accelerated computing. Nvidia has more than 80% of the GPU networking market and AMD is catching up; as agentic AI becomes the next growth driver, AMD has a unique edge because it is one of the two big players in both desktop and server CPUs; Broadcom has built TPUs for Google for years and now has opportunities with big players like OpenAI; TSMC fabs for all three, so it wins no matter who wins, an almost oligopolistic or even monopolistic position.
— Glen KacherOpen-source models are fighting a price war
Kacher thinks the battle between Anthropic and OpenAI is happening in real time, and that Google remains a player thanks to Gemini and its search distribution advantage, so it cannot be ruled out. But the battlefield really emerging is open-source models: they are free, can be downloaded to run on local hardware, or run on someone else's commoditized hardware in the cloud, competing directly with the two big companies' expensive frontier models. He judges that early signals show both approaches have their place. He also points to the regulatory problem: open-source models can be put inside your own software and tuned to your needs, so regulators actually have few options, because these models are already out in the wild.
— Glen KacherPeople were talking about GPUs for AI back in 2005
Kacher recalls that while at Integral, Kleiner Perkins partner Bill Joy mentioned a group of engineers at Caltech using GPUs for early AI computation, around 2005 or 2006, and Bill Joy's conclusion was that GPUs would be the chip architecture best suited to AI computation. It stayed in his head. Later, on every visit to Nvidia, Jensen would talk about AI, but at the time AI was only a tiny end market for Nvidia, and crypto mattered far more. In the second half of 2022 the crypto crash and the takeoff of AI happened at the same time, the market was more focused on crypto and pushed the stock down, which gave them the chance to build a position in Nvidia.
— Glen KacherThe giants are not over-investing, they are catching up
Faced with the challenge that AI could be over-provisioned like fiber, Kacher's answer is: Amazon, Microsoft and Google are building the compute stack together with Anthropic and OpenAI, and the question is how far ahead of demand they are planning. His judgment is that today they are not ahead, they are behind and catching up. Bears expect them to over-invest, but that is not happening today. He concedes there are real bottlenecks, including the memory companies and TSMC, which cannot invest as fast as they would like, so demand today still runs far ahead of supply, and the pessimism around AI is misplaced.
— Glen KacherA 15-year cycle, and only a third of the way in
Kacher likens AI to the shift from client-server to internet architecture, a computing cycle that happens roughly every 15 to 25 years. The old technology was a search-and-retrieve model, with data sitting in databases; AI generates a custom answer to your question and your data every time it is used, and is far more compute-intensive. The rhythm he gives: infrastructure development takes roughly 5 to 10 years, years 6 to 16 are when the platform or operating system takes shape, and years 11 to 21 are when applications become the main venue for enterprise investment and innovation. By that framework, we are about a third of the way through the first phase, and the whole cycle is at least 15 to 20 years.
— Glen KacherPower and local politics have become the new bottleneck
Kacher treats political resistance to data centers as a real risk: any bottleneck that slows technology adoption is a problem. He says it is not yet a big problem today, but it is an emerging one. His solution is that the industry has to explain itself proactively: Northern Virginia, as the data center capital of the world, has brought the area tax revenue and blue-collar demand for electricians, plumbers and construction workers, a shot in the arm for the economy. He mentions a data center project in San Mateo that plans to put Bloom Energy natural gas fuel cells behind the meter, with almost no emissions and almost no noise, and residents still opposed it because they had "heard data centers are bad." He sees this as an education problem, requiring project-by-project explanation of energy, jobs and tax revenue, from local to national.
— Glen KacherIn their own words · checked verbatim
the number one player is going to get two-thirds of the market. number two player might get 20%, 25% tops, and everyone else fights for the scraps
Glen Kacher7:16
I think people right now are, for instance, underestimating NVIDIA's opportunity to innovate as well.
Glen Kacher9:21
The negative doomers are expecting them to over-invest, but today that's just not happening.
Glen Kacher45:36
So today, demand is still running way ahead of supply. And so this doomerism that has grown up around AI, in my mind, is misplaced.
Glen Kacher46:39
If we follow history, these things tend to to 15 years to become a quarter of the total capacity of the industry. So to say that it's going to take multiple decades is not much of a stretch.
Glen Kacher51:47
It leads to users consuming 5X the tokens that you would consume just directing AI as you would a search engine.
Glen Kacher1:02:01
you look at what's happening today with AI and people taking this incredibly powerful technology that is going to change the world and is already starting to change it, and making it this evil empire.
Glen Kacher1:05:05
Figures
| Nvidia's share of the AI accelerator market | about 85% | 8:17 |
| Nvidia's share of the GPU networking market | more than 80% | 21:51 |
| TSMC, Nvidia, Broadcom and AMD as a share of public holdings | about 40% | 21:51 |
| Number of private deals Kacher made in 13 years at Integral Capital | 46 | 6:12 |
| Light Street's 2021 return | down 26% | 34:15 |
| Light Street's 2022 return | down 54% | 34:15 |
| Light Street's 2023 return | up 46% | 34:15 |
| Light Street's 2024 return | up 59% | 34:15 |
| Light Street's 2025 return | up 37% | 34:15 |
Glossary
- XPU
- An umbrella term that swaps the G in GPU for an X, referring to the many kinds of AI chips used for inference and other workloads.
- agentic AI
- A form of AI that keeps working on a problem even when you are not directing it.
- SaaS-pocalypse
- The market view that AI will wipe out traditional software companies en masse.
- behind the meter
- Power that does not pass through the public grid's metering, generated on site by the data center and supplied directly.
How to listen
Investors watching AI compute, semiconductors and tech stocks, and founders who want to understand the power and local-politics risks around data centers.
The opening 0:00-14:35, covering his résumé and lineage, can be fast-forwarded; low information density.