The IPO Comeback: Why Tech Giants Are Finally Going Public | All-In Liquidity IPO Panel

Why does it matter? Because going public changes almost nothing — and most of the money gets made by whoever holds the shares longest

Two founders compare notes on the morning after. Andrew Feldman rang the Cerebras bell three weeks ago. Will Marshall watched Planet Labs run from $5 to $50 in 12 months. Together with investor Brad Gerstner, they make the case that IPO timing is the wrong obsession — and explain what's actually driving the two biggest secular shifts in compute: radical silicon architecture and solar-powered orbital data centers.

• Cerebras is 15-18x faster than GPUs for OpenAI because it solved the actual bottleneck — memory-to-compute bandwidth — with a chip the size of a dinner plate. • Space data centers cross economic parity with terrestrial ones at $200-300/kg launch cost, a threshold Starship reaches in roughly 2 years. • LLMs trained on internet text are structurally blind to physical reality; satellite earth observation is the missing sensory layer. • Historically, more VC returns are generated after IPO than before — but LP pressure to distribute at the bell keeps giving those gains to public market investors.

If your chip looks like Nvidia's, you've already lost

Cerebras runs inference 15 to 18 times faster than a GPU for OpenAI — because Feldman decided in 2015 that copying Nvidia's architecture was a guaranteed path to irrelevance.

"If you want to be 20 times better than somebody, your architecture can't look like them," he says. "They have enjoyed and eaten all the low-hanging fruit. If you build a GPU, the odds that you're better than Nvidia are approximately zero."

The bottleneck Cerebras went after wasn't raw processing speed — it was the movement of data from memory to compute. "This is the fundamental problem in AI." GPUs are designed around that constraint; they've optimized within it for decades. Cerebras decided to eliminate it. The answer: a chip the size of a dinner plate, with memory sitting directly next to compute and using a faster memory type that postage-stamp-sized chips structurally can't support.

The speed advantage follows directly from that one architectural bet. Feldman frames why it matters through a counterfactual: how big is the market for slow search today? Zero. How long do you wait for a website before clicking away — three seconds, five? "You will not wait for AI. We have to deliver it to you in real time."

The framework travels beyond chips. When attacking a dominant incumbent, the question isn't how to be incrementally better at what they do. It's which fundamental constraint they've accepted that you could eliminate entirely. Nvidia treated the memory bottleneck as a given. Cerebras treated it as the problem.

$200-300 per kilogram — the exact threshold that flips space data centers from crazy to inevitable

Within 2 to 3 years, it will be literally cheaper to run a data center in orbit than on the ground. This isn't a vision statement — it's a calculation Planet Labs ran with Google nearly a decade ago.

The math: at today's ~$1,000/kg launch cost, orbital compute doesn't pencil. Cross $200-300/kg and it does, decisively. Launch costs have already dropped 10x over the past decade. Starship's trajectory puts the threshold within reach shortly. Marshall is careful: "Elon might say it's next week, but realistically a couple of years."

The energy economics seal the case. Data centers are fundamentally a power problem. Solar is the cheapest watt available, but terrestrial solar is intermittent — so you bolt on batteries or gas backup and cost climbs steeply. In a sun-synchronous dawn-dusk orbit, a solar panel faces the sun continuously, 24/7, generating five times more energy per panel than ground-based solar. No batteries. No backup. Just panels, chips, and RF signals.

Planet Labs has already launched Nvidia GPUs and is testing Google TPUs in orbit. Feldman flags what remains hard: "Building the clusters in space requires inter-chip communication — and we're not even good at doing that on the ground." He frames it as potentially a self-driving-style "last 10% is 80% of the time" problem. Marshall's counter: get launch costs low enough to run cheap experiments, and the engineering finds its solutions. The threshold doesn't just make space cheaper — it starts the clock on that work.

AI didn't make computers faster — it unlocked domains they'd been shut out of for 70 years

The AI compute boom looks like an upgrade cycle. Feldman's framing resets that entirely.

"We were bad at images for almost the entire history of compute. We could store them and that's about it. We were bad at language. We could store it but that's about it." Numbers, computers mastered. Everything else — foreclosed. What shifted around 2015-16 wasn't performance. It was domain access: computers could suddenly extract meaning from images, generate language, understand rather than merely retrieve. Entire categories of knowledge previously inaccessible became addressable.

At the same time, the world was producing vastly more images — from phones, terrestrial cameras, satellites. New supply and new capability arrived together. "This is what's underpinning Nvidia's growth and all the growth you're hearing about in AI compute," Feldman explains. Hardware builders "suddenly could attack more and different parts of knowledge."

The distinction matters for how you model demand durability. An upgrade cycle implies: buy the new chips, settle into steady state. But if AI unlocked domains that previously generated zero compute demand, the addressable market is permanently larger. Every image ever taken, every sentence ever written, every sensor reading ever logged — all newly valuable as training data and inference target. That's not a cycle. It's a structural expansion of what computers can be hired to do, and it doesn't peak when the current generation of models is deployed.

LLMs have absorbed the entire internet — and have no idea what's happening on the ground

Ask a current LLM what's growing in a specific field in Kansas right now, or whether a particular refinery is operating, or where troops are massing on a border. The model has no answer — not because it's undertrained, but because that information was never on the internet.

Marshall's diagnosis is direct: "All the cool stuff we're doing with LLMs now is really based on just the text of the internet being absorbed into these models — which is incredibly powerful already — but they don't know about the real world. They don't know about that farm field, that flood, that security situation around the corner."

Planet's answer: "large earth models" — LLMs grounded in daily, time-series satellite imagery of the entire planet. "If you give them real world data, then they can answer real world problems. And that's going to open up gazillions of applications." Marshall's framing extends further: not just AI but "planetary intelligence" — planetary sensing systems in space feeding planetary compute systems in space.

The moat here isn't model architecture, which will commoditize. It's data provenance: 200 satellites imaging the full Earth every day, a time series going back years, updating continuously. That data doesn't exist elsewhere. As text-trained LLMs converge toward commodity, the compounding advantage migrates to whoever owns unique physical-world data streams that the internet simply never captured.

Planet went from $5 to $50 after the IPO — and the investors who demanded distribution at the bell missed all of it

The investors who stayed with Planet Labs through its post-IPO run captured 10x. Gerstner's fund, a decade earlier, didn't stay in a different company — and the aftermath still stings.

"Historically more money's made after IPO than before," Feldman says flatly. "Every single study shows there is more money to be made both in percentage and in absolute." The structural logic: pre-IPO, the capital a fund can deploy in any single company is constrained by ownership and round size. Public markets can absorb orders of magnitude more. When a company keeps executing, the absolute gains dwarf what early checks could access.

Gerstner's specific case: invested at a $1 billion valuation, distributed shares at $3-4 billion — then watched the stock run to $50 billion over the following 24 months. LPs called asking why he didn't hold. His answer: "Because you were pounding on us to distribute the shares."

At Planet Labs, Google hasn't sold a share. Capricorn held until recently. The investors who stayed got the full move. Those who demanded distribution handed a 10x to whoever bought their shares in the open market.

Cerebras's structural response: a "dribble lockup" releasing shares incrementally over six months against performance hurdles rather than one cliff event. SpaceX is reportedly structuring something similar. The forced-distribution problem is a mechanism design problem — if the lockup cliff creates the perverse incentive, change the cliff.

The day after the IPO: no new customers, no engineering progress, same bad vendor relationships

The day after Cerebras rang the bell, it had sold no more product than the day before. Engineering had made zero progress. Bad vendor relationships were still bad.

Feldman's summary is the most useful 30 seconds in the conversation: "You have this enormous event and the next morning you've sold no more stuff. Your engineering projects have made no progress since the day you weren't public. If your relationships with your vendors are bad, they're still bad. Not a damn thing changes in the important parts of your business."

What the IPO does deliver: more cash, and a durability signal that closes certain enterprise deals. Marshall's point on this is concrete — Planet's government and agricultural customers need to know the company will still exist in five years. Being publicly listed is a credibility anchor private companies can't replicate. "They don't want you to just disappear."

What it doesn't deliver: operational improvements, accelerated product development, or resolution of structural weaknesses. Go public when the liquidity and legitimacy benefits outweigh compliance costs — not because you expect it to fix what's broken.

Everyone's watching launch costs. The bigger story is that a billion-dollar satellite now fits in a backpack

Cheaper rockets get all the attention. Marshall argues satellite miniaturization is actually the underappreciated driver — and the two vectors together are what make the current moment genuinely structural.

"A thing that people don't know that is actually perhaps more important is that we've had a miniaturization of satellites. So that the same satellite that used to cost a billion dollars and weigh 20 tons now costs a few kg." Same capability, orders of magnitude cheaper. "It's the same as the mainframe computer to desktop computer revolution for space — and it's unlocking loads of applications."

Launch costs down 4-5x over the past decade. Satellite hardware costs down by far more. Together, those two shifts made Planet's 200-satellite constellation economically feasible — a daily-imaging capability that was theoretically possible for decades but structurally out of reach until both vectors moved.

The frame for evaluating any new space venture: look at both simultaneously. A company that benefits from cheap launch but still requires expensive custom hardware hasn't really solved the problem. The compounding effect only kicks in when miniaturization and launch economics converge.

When sensing, compute, and power all converge in orbit, there's no terrestrial equivalent to build

Marshall almost said it out loud: "planetary intelligence" running on "planetary compute systems in space." Follow that thread to its end.

Once large earth models run on orbital compute that's simultaneously collecting their training data — powered by continuous solar, launched below $300/kg, on hardware weighing kilograms instead of tons — the gap between observation and insight collapses to zero. Nothing on the ground can replicate that loop. The companies that own both the sensing and compute layers in orbit won't be selling data subscriptions. They'll be operating the only infrastructure in existence that can answer questions the internet never even knew to ask.

---

Topics: IPO, AI hardware, Cerebras, Planet Labs, space compute, satellite, earth observation, venture capital, LPs, Nvidia, chip architecture, space-based data centers, LLMs, public markets, lockup