Tech Founders Argue IPOs Can Create More Upside After Listing

Brad Gerstner put the investor question directly: when a company goes public, should venture investors distribute shares immediately, or hold through public-market value creation? His framing was aimed at limited partners who want liquidity but may also miss upside if they force distribution too soon.

Planet Labs was his counterexample to immediate distribution. The company went public in 2021 through a SPAC at about $2 billion, according to Gerstner. He said most of the value was created in year three or four after the listing. Will Marshall said most early investors stayed in and captured that upside. Google, he added, had not sold a share and remained Planet’s largest single investor; Capricorn had not sold until very recently. Marshall’s view was that those investors were smart to hold and should hold even more.

Many LPs, Gerstner said, want shares distributed once a company goes public. He compared Planet to Altimeter’s experience with MongoDB: Altimeter invested pre-IPO at about $1 billion, distributed shares around $3 billion or $4 billion, and then the company went to $50 billion over the following 24 months. Investors later asked why Altimeter had not held the shares, even though they had been pressuring the firm to distribute them.

Andrew Feldman’s view was categorical: more money is historically made after the IPO than before. He said studies show this in both percentage terms and absolute dollars. The reason, as he described it, is that venture investors can usually put only modest amounts of money to work before a company goes public. By the time the company is public, there is more capital in the business, performance is clearer, and the opportunity to make substantially more can be larger.

Cerebras also used what Gerstner called a “dribble lockup,” in which shares can be released gradually over six months according to “a bunch of performance hurdles.” Gerstner described it as an innovation with the banks and said SpaceX would likely have a similar lockup structure. The panel did not detail the specific performance hurdles.

The largest private companies pose a different version of the same problem. Marshall noted that most big tech companies went public at valuations of a few billion dollars, not a few trillion, leaving enormous upside for public investors. If SpaceX were to provide a comparable post-IPO “liftoff” from a multi-trillion-dollar starting point, he said, investors would have to believe in quadrillion-dollar valuations.

Gerstner argued that Anthropic, OpenAI, and SpaceX should not necessarily be treated as the new normal. He said public markets may be shifting back toward earlier listings, with portfolio companies considering IPOs at $1 billion, $3 billion, or $5 billion. He contrasted that with a decade in which staying private for as long as possible was heavily promoted. Planet, in his telling, is an example of venture-style value creation happening in the public markets rather than being captured entirely by private investors.

Chamath Palihapitiya added a governance and performance argument. Getting public sooner, accepting public-market scrutiny, and being forced to deliver can sharpen a company. His phrase was that “steel sharpens steel, iron sharpens iron.” Public markets, in that view, are not merely a liquidity venue. They are a discipline mechanism that may improve execution while letting more investors participate in the upside.

For Andrew Feldman, the newly public status of Cerebras did not instantly transform the company’s operating reality. He described the IPO process as full of “garbage”: large Zoom meetings, repeated document reviews, small edits that move commas without adding value. The public-market event itself was enormous, but the next morning, he said, the company had not sold more product and its engineering projects had not advanced simply because the listing had happened.

The practical change, in Feldman’s telling, is that the company has more money in the bank and a new set of constituents to communicate with. The fundamentals are unchanged. If a company needed new supply before the IPO, it still needs new supply afterward. If vendor relationships were bad, they remain bad; if they were good, they remain good. Employees celebrate, the company receives external validation, and then the work resumes.

Brad Gerstner supplied the investor-side context around Cerebras’s route to market. He said Cerebras had tried to go public earlier, that one investor was the UAE, and that questions around CFIUS under the Biden administration made the path challenging. His characterization was that “everything was really hard until it got really easy”: roughly nine and a half years of difficulty followed by a period in which demand for the IPO accelerated. Gerstner said the company priced at $185 after the range had been raised twice, opened around $320 per share, and was trading around $230, implying a market capitalization in the $50 billion to $60 billion range.

The companies’ market debuts were presented through Nasdaq and NYSE imagery: Cerebras with celebration footage and chip imagery, and Planet with the NYSE opening bell on December 8, 2021. Feldman’s own explanation of timing was less triumphant. Asked how Cerebras got the timing right, he answered that it did so “by getting it wrong for a decade.”

The IPO mattered most visibly to the people who had endured that decade. Cerebras brought employees who had been with the company more than nine years, along with their families. Feldman said he learned that engineers owned ties and could survive wearing them. More seriously, he was struck by how much the milestone meant to families, including parents who could finally see the company recognized in a public way. One Chinese immigrant father of a Cerebras leader told him, “I thought it would have happened faster.”

Will Marshall framed the public-company transition similarly, but with more emphasis on customer credibility. Planet Labs went public in 2021, and Marshall said the listing gave liquidity to early employees and investors, provided cash to the company, and helped customers believe Planet would endure. That matters for a business serving large agricultural companies, governments, defense and intelligence customers, and countries that depend on Planet’s data. Public-company status, in his words, gives a company “force in the world”: it signals access to capital and staying power.

Marshall said Planet does not focus on the daily stock price. He credited the market with starting to understand the importance of space infrastructure and data, but his operating emphasis remained long-term shareholder value. The shared point from both CEOs was narrow but important: public markets add capital, legitimacy, scrutiny, and liquidity, but they do not substitute for execution.

David Sacks pressed Marshall on whether Planet’s market framing had changed. The question was whether Planet had moved from being seen as a data source for people who needed satellite imagery and maps into something closer to a military-technology company, especially after the success of companies such as Anduril.

Marshall resisted a simple reclassification. Planet, he explained, operates the largest Earth-imaging fleet, about 200 satellites, imaging the entire Earth every day. He compared the product to the satellite layer in Google Maps, except current rather than years old, with a daily historical archive that enables time-series analysis of activity on Earth. The use cases include agriculture, energy, civil governments responding to floods and fires, and security.

The revenue mix nevertheless gives defense and security unusual weight. When Sacks asked what percentage of revenue or customer base was military, Marshall said security accounts for about 60% of Planet’s revenue today. Security was part of the company’s plan from the beginning, but it has become a larger fraction than he might have guessed. His explanation was geopolitical: current conditions demand earlier visibility into threats.

Planet’s data, Marshall argued, can let customers see “threats round the corner,” with weeks or months of advance warning. He presented that as conflict-preventing rather than merely conflict-serving: better information can allow actors to stop conflict before it starts. Asked directly whether he was reticent to be perceived as a military company, Marshall said no, but added that Planet should not be limited to that perception. The company also works with farmers, energy companies, civil governments, NASA, and others.

The broader change is that space capabilities once reserved for governments are becoming accessible to commercial and civil users. Rocket launch costs have fallen sharply, but Marshall said a less appreciated shift is satellite miniaturization. A satellite that once might have cost a billion dollars and weighed 20 tons can now weigh kilograms, tens of kilograms, or hundreds of kilograms and do as much or more. He compared this to the move from mainframes to desktops: lower cost and smaller form factor unlock a wider set of applications.

Will Marshall’s strongest claim about Planet was not merely that it collects imagery, but that AI lowers the barrier to using that imagery. The company’s satellites generate a daily, global, historical data layer; AI makes it easier for more customers to turn that layer into decisions.

He estimated the market for Earth observation and AI applications at $75 billion to $100 billion. The near-term opportunity is applying large language models to Earth imagery data, opening uses in agriculture, energy, civil government, permitting, and other domains. The value of the AI depends on the data it can see. Text-trained models may be powerful, but Marshall argued they are “blind” to the physical world unless connected to real-world observation.

That line clarifies why Marshall sees space and AI as converging. Satellite systems provide planetary sensing; AI systems provide interpretation and access. He described possible future systems as “large earth models” rather than large language models, or “planetary intelligence” rather than simply AI. In that framing, Planet’s daily Earth-imaging archive matters because real-world data lets AI answer real-world problems: a farm field, a flood, or a security situation “around the corner.”

Marshall’s analogy to Google was explicit: just as Google indexed the internet and made it searchable, Planet is indexing the Earth and making it searchable. The source reinforced that claim with a Planet satellite-imagery interface showing search results beside a map of Damascus, Syria, with dated SkySat and PlanetScope imagery. The difference from a conventional web index is that Planet’s index is physical and continuously refreshed. AI, in this view, is the interface that can make that index useful to a broader set of customers.

Chamath Palihapitiya pushed Marshall to explain space-based data centers: whether they are viable, what they are, and how they would work. Marshall’s answer centered on cost curves and power.

Planet and Google studied the economics eight or nine years ago, Marshall said, comparing ground-based data-center costs with the costs of putting data centers in space. Their conclusion was that when launch costs fall to roughly $200 to $300 per kilogram, space-based data centers become cheaper. He said launch costs today are just over $1,000 per kilogram, down about 10x in the last decade, and that on the current trajectory, particularly with Starship, he expects the $200 to $300 range within two to three years. He added that Elon Musk might say it is next week, but Marshall’s own estimate was “a couple of years.”

The power argument is straightforward in Marshall’s telling. Data centers are a power problem. Solar panels are the cheapest way to get a watt, but terrestrial solar is intermittent, requiring batteries, gas, or nuclear backup, which increases cost. In a sun-synchronous dawn-dusk orbit, a solar panel can face the sun around the clock. Marshall said such a panel can gather five times more energy than a comparable panel on the ground, without batteries or other backup infrastructure. A space compute system, in his simplified description, becomes solar panels, chips, and RF signals up and down.

Planet is already testing pieces of that future. Marshall said the company has launched some Nvidia GPUs into space and is partnering with Google to launch TPUs into space for an early test. He acknowledged that much technology remains to be figured out, but argued there is “no question” that within 10 years most compute will be put in space. He said the market would be measured in trillions and would be larger than current space businesses such as communications or Earth imaging.

Andrew Feldman was more cautious. He agreed that the problem is important and worth attacking, but said “one or two hard problems” remain beyond putting GPUs in space. The central difficulty is building the clusters needed for communication among chips or systems. Sacks interjected that the industry is not good at doing that on the ground; Feldman agreed and added that it is “really not good” at doing it in space.

The contrast was Marshall’s cost-curve confidence against Feldman’s cluster-engineering caution. Feldman said space compute certainly will occur, but he has it on a different timeframe. He compared the risk to self-driving cars, where the last 10% of the problem consumed a decade of work. Lower launch costs are the condition that makes experimentation possible: once launches are cheap enough to permit failed experiments and iteration, the engineering questions can move from paper into practice.

Andrew Feldman’s account of Cerebras begins with a broader claim about the history of computing. For most of that history, he said, computers were excellent with numbers and poor with images and language. They could store images and text, but not do much with them. AI changed that around 2015 and 2016 by opening problems that had previously been hard for computers to address: finding insight in images, generating language, and understanding language rather than merely storing and regurgitating it.

That change created the demand behind the growth in AI compute. As more images and language became addressable by computation, processor builders suddenly had a larger universe of knowledge to attack. Feldman placed Nvidia’s growth and the broader AI compute boom inside that shift.

Cerebras’s founding bet was that AI would consume enormous amounts of compute and that new workloads create openings for market share to move. Feldman offered several historical examples. Graphics created the dedicated GPU and Nvidia. Cellphone compute shifted share away from Intel and AMD toward ARM. Data networking in the late 1990s created openings for companies such as Cisco, Juniper, and Arista. AI, in his view, presented the same kind of architectural discontinuity.

Cerebras made two bets: dedicated silicon would be the answer, and it could not look like a GPU. Feldman said that if a company wants to be 20 times better than an incumbent, its architecture cannot resemble the incumbent’s, because the incumbent has already captured the obvious gains. In his view, building another GPU and expecting to beat Nvidia offered approximately zero odds of success.

The technical problem he emphasized is moving data from memory to compute. Cerebras’s answer was to build a chip the size of a dinner plate, rather than the size of a postage stamp, and put memory close to compute. The larger chip allowed use of a faster type of memory, which opened performance gains. Feldman said that when OpenAI uses Cerebras, it is 15 to 18 times faster than a GPU.

The business importance of that speed is latency. Faster inference means answers arrive more quickly, the AI experience becomes more enjoyable, and users can apply AI to harder problems without waiting. He framed the issue through familiar counterfactuals: there is no market for slow search, no market for dial-up, and users will not wait long for a website to resolve before abandoning it. AI, he argued, must be delivered in real time.