AI compute will be exhausted, but its owners probably won't be
In the last fiber-optic bubble, the most-indebted companies collapsed first, but the fiber assets kept working when others took them over. Mike Volpi believes this GPU buildout will follow the same script: the assets will be fully utilized, but their ownership may well change hands.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Bandwidth jumped by technology breakthroughs; compute is stuck drinking from Moore's Law
In the 1990s, fiber-optic bandwidth multiplied 40-80x through WDM (wavelength-division multiplexing)—packing multiple colors of light down the same fiber. AI compute expansion is fundamentally different: constrained by Moore's Law, it grows slower. Volpi sees this as the root technical difference between the two buildouts—the pace of supply growth is not even close.
— Mike VolpiAssets will be exhausted; their owners will likely not
Volpi reviews the Internet bubble: companies most burdened by debt collapsed first, but their fiber assets didn't disappear—others picked them up cheap and kept using them. He expects the same script to play out: data centers and chips will be fully utilized, even beyond expectations, but whoever ends up owning these assets may well not be the companies building them today.
— Mike VolpiH100s should be obsolete; their rental price keeps rising
Four years ago, conventional wisdom said H100-class chips had about five years of useful life—and rented for roughly $1.82 per GPU-hour. By all rights they should have cheapened as they neared end-of-life, but instead they now rent for $2.50-3.00 per hour. Volpi attributes this to extreme compute scarcity: as long as semiconductor production lags demand, older chips stay needed, and asset lifecycles stretch far beyond what finance models assumed.
— Mike VolpiCisco's M&A philosophy: accept one-third total loss
Volpi recalls Cisco's acquisition method: split outcomes roughly into thirds—one-third worked great, one-third was mediocre, one-third failed outright. The company tolerated the failure rate as long as it did it often enough to hit several genuine home runs. Some acquisitions cost $500 million and generated $1 billion in profit within two years. By around 2000, half of Cisco's revenue came from acquired businesses, and no single acquisition ever exceeded 4-5% of the company's market cap.
— Mike VolpiNVIDIA's spending on neocloud startups will fail to reshape the market
On NVIDIA's practice of investing in numerous neocloud companies to cultivate more buyers, Volpi argues it won't change the industry's final shape: in any sector where scale dominates, no amount of money pumped into small players will stop consolidation into a handful of giants—like how US aviation ended with three major carriers. NVIDIA does this not for financial returns but to avoid negotiating with only three or four mega-customers, but it's only delaying, not altering, the industry's natural evolution.
— Mike VolpiModel routing emerges as token prices diverge 5-7x
For the past two years, the market default was to use only the most advanced frontier-model tokens. But Volpi observes the market growing smarter: top-tier Anthropic tokens and top-tier Kimi tokens differ by 5x, even 7x in price—a gap wide enough to spawn a new business called ‘model routers’ that split requests to different-priced models by task difficulty. He sees this making AI's total demand price-sensitive for the first time, instead of blindly chasing the most expensive model.
— Mike VolpiOrder bubbles burst: Cisco's market cap fell to one-tenth in six months
Volpi recounts Cisco's collapse: tight delivery windows led customers to habitually over-order to guard against allocation cuts, and Cisco transferred risk to suppliers by paying early to hoard stock. When true demand softened, orders dropped directly to one-quarter of what they'd been. Combined with risk already locked into the supply chain, the company missed 35-40% of quarterly revenue in a single quarter. Stock fell from over $80 to $9, a ninefold decline in six months.
— Mike VolpiAI data centers should abandon stat-muxing for deterministic circuits
Internet-era switch design assumed bursty traffic, using stat-muxing to smooth peaks and valleys. AI data centers run GPUs at full load constantly, with highly predictable, nearly-deterministic traffic—especially in training. Volpi argues that today's mainstream Broadcom Tomahawk stat-muxing architecture is ‘enough bandwidth but wrong shape’—what should exist is a new switching architecture designed for full-time load, needing no buffers. This is classic innovator's dilemma: incumbent vendors won't cannibalize their own business. The opportunity goes to new entrants willing to all-in on the new architecture.
— Mike VolpiIn their own words · checked verbatim
So that's a long-winded answer to say, do I think the build-out that's happening now is useful and will be fully utilized, if not more? 100%. Will it be owned by the same people that own it today?
Mike Volpi6:14
Not at all defunct. No, not only that. It rents at a higher rate.
Mike Volpi11:33
we were not afraid to fail, and we did it with high enough frequency that we got a few outliers
Mike Volpi14:43
No matter how much you finance them, the little ones are always weaklings.
Mike Volpi19:54
right now we live in a world where demand for AI capacity, tokens, whatever, is near infinite. And supply is incredibly constrained.
Mike Volpi33:08
the economic difference between a top end anthropic token and a top end Kimmy token is like five to one or seven to one or whatever. It's a massive economic difference.
Mike Volpi34:09
Stock went from 80 some dollars a share to nine dollars a share, I want to say. So it was like literally one tenth the value over a period of six months.
Mike Volpi45:36
Figures
| NVIDIA-counted neocloud companies | approximately 200 | 7:18 |
| Cisco acquisitions completed | approximately 70 | 18:54 |
| Cisco per-deal limit | no more than 4-5% of company market cap | 15:49 |
| Cisco stock during collapse | from $80+ to $9 | 45:36 |
| Cisco quarterly revenue miss during collapse | 35-40% sequential decline | 44:31 |
| Price spread between top-tier Anthropic and Kimi tokens | 5-7x | 34:09 |
| AI infrastructure investment as % of GDP | approximately 2% | 7:18 |
Glossary
- WDM
- Wavelength-division multiplexing; packing multiple signal colors down a single fiber to multiply bandwidth.
- Stat-muxing
- Statistical multiplexing; traditional switching that shares bandwidth by smoothing traffic bursts.
- Neocloud
- Emerging GPU cloud providers that rent compute capacity.
- KV cache
- Key-value cache storing intermediate attention results in GPU memory during inference.
- Model router
- A system routing requests to different-priced models based on task difficulty.
How to listen
Investors, founders, and engineers interested in AI infrastructure investment cycles, GPU asset pricing and M&A strategy, and network architecture.
The introduction of Mike Volpi's background and opening remarks (0:00-3:00) can be skipped without losing the main argument.