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

This source is an authorized transcript of a live/podcast-style technology and investing discussion focused on public markets, Cerebras, Planet Labs, AI silicon, satellite data, space-based data centers, and the distribution of value between private and public investors.

The transcript explicitly introduces Will Marshall of Planet Labs and an Andrew Feldman as the Cerebras participant, but the automated transcription renders Andrew’s surname variously as “Bubman,” “Feland,” and “Feldman.” It similarly renders Brad Gerstner as “Brad Gersner/Gersonner” and his firm as “Autoimmet Capital.” These names should be independently checked before publication rather than silently corrected.

At least one host/moderator is identifiable from context as Jason Calacanis, because he recounts seeing a security-screen entry reading “Jason McCabe Calacanis with Donald J. Trump.” Other moderator/interviewer turns are not consistently distinguishable from the transcript alone and are therefore attributed below as Host/Moderator when identity cannot safely be established.

• The transcript contains severe rolling-caption duplication: most phrases occur two or three times.
• Speaker labels are largely absent.
• Several proper names and company names are clearly mistranscribed.
• Some numerical phrases run together, including “5060 billion dollars,” which could mean several different things and is therefore preserved as [transcript unclear].
• “CPHAS” appears in discussion of UAE investment and regulatory difficulties; the intended term is [transcript unclear].
• “spa” is used for Planet Labs’ route to the public market and should be checked; the intended acronym may have been mistranscribed.
• Several investor names are badly garbled.
• A few sentence transitions are incomplete because speakers interrupt one another.
• Nothing in this Stage 1 record independently verifies the speakers’ factual claims, estimates, predictions, investment assertions, or technical performance claims.

The opening montage frames 2026 as potentially an exceptional year for IPOs and describes Cerebras as the major AI IPO of the year so far.

The Cerebras executive says the company is participating in something extraordinary and claims that, in what it does, Cerebras is “the fastest bar none.”

Will Marshall is introduced as co-founder and CEO of Planet Labs. Marshall frames space and AI as converging technologies, saying they are effectively “getting married.” His analogy is that Google indexed the internet and made it searchable, while Planet is indexing the Earth and making it searchable.

Brad [surname transcription varies] is introduced as the founder and CEO of [firm name transcript unclear] Capital, described as a leading technology investment firm.

Brad argues that the current technology wave is the largest in the history of technology and predicts that it will be extremely beneficial for the United States. He says his support for the companies involved comes from wanting America to succeed.

Before the core panel discussion, a host recounts an anecdote about Davos.

He says he had intended to take his annual ski trip to Japan when “Sax”/“Sacks” called him. There is joking reference to “Tucker.”

According to the host:

• he cancelled his previous plans after the call;
• Sacks told him POTUS needed him in Davos as a moderator;
• he was given only three days’ notice;
• he agreed and travelled to Davos;
• he received a special green security badge;
• while being cleared through security, he saw a monitor displaying the name “Jason McCabe Calacanis with Donald J. Trump”;
• he found this hilarious;
• he subsequently participated in an interview there;
• he says the group conducted approximately six or seven All-In interviews during the trip.

The discussion then transitions away from the anecdote and toward the newly public companies represented onstage.

The host frames Cerebras and Planet Labs as gateways into two major technology trends:

1. AI silicon
2. space infrastructure and space-based data centers

The panel first pauses those subjects to discuss what actually happens when a company becomes public.

The Cerebras executive had been public for only approximately three weeks, according to the discussion. Will Marshall had been dealing with the public markets considerably longer.

Andrew describes the IPO process itself as extraordinarily burdensome.

He says there is a huge amount of administrative work that adds little obvious value to the underlying company. Examples include:

• Zoom meetings with as many as approximately 130 participants;
• repeated reviews of documents;
• minute edits in which little more than punctuation appears to change;
• considerable executive and organizational time consumed by the process.

His broader point is that the IPO ceremony can feel enormous while the actual operating business wakes up the next morning fundamentally unchanged.

He notes:

• the company has not automatically sold more products;
• engineering projects have not suddenly advanced;
• supply problems do not disappear;
• poor vendor relationships remain poor;
• strong vendor relationships remain strong;
• the company does have more money in the bank;
• management gains new constituencies, particularly public shareholders and others to whom it must communicate.

His description of the aftermath is essentially: employees celebrate, everyone is excited, people high-five — and then everyone goes back to work.

He therefore distinguishes capital-market transformation from operating transformation. Going public can change financing and stakeholders without magically repairing product, engineering, supply, sales, or execution problems.

Another panelist or moderator, speaking as an investor and former board member, adds context about Cerebras.

He says his firm had invested in the company and that he had been on its board during an earlier effort to go public.

He characterizes Andrew as someone who had survived roughly a decade of difficult corporate challenges.

He says Cerebras faced an unusually difficult public-market path because one of its investors was from the UAE, creating questions involving “CPHAS” [transcript unclear]. He associates the regulatory problem with the prior Biden administration and says it made the effort to become public more difficult.

His outside observation is that Cerebras experienced approximately:

• nine and a half years of extreme difficulty, followed by
• roughly 12 months in which conditions became much easier and investors suddenly wanted exposure.

He presents the trajectory as a case where something can be difficult for years before sentiment and circumstances change rapidly.

The speaker states that:

• the IPO was priced at 185;
• the proposed price range had been increased twice;
• the stock opened at approximately $320 per share, though he qualifies this with “I think”;
• at the time of the conversation it was around $230 per share;
• he then gives a market-cap figure transcribed as “5060 billion dollars” [transcript unclear].

These figures require verification before publication.

The conversation asks whether the IPO had produced softer benefits:

• increased employee morale;
• increased company profile;
• less distraction after the event;
• more available capital;
• potentially greater credibility with enterprise customers.

Andrew answers by saying people frequently ask how Cerebras managed to get its IPO timing right.

His answer is deliberately paradoxical: they got the timing right by getting it wrong for a decade.

His point is that the successful timing should not be interpreted as a single brilliant market-timing decision. It followed more than ten years of work and previous unsuccessful or mistimed attempts.

Cerebras brought employees who had been with the business for more than nine years to participate in the IPO celebration and invited their families.

Andrew jokes that he discovered engineers owned ties and could wear them without dying.

More substantively, he says he underestimated how emotionally important the event would be to employees and their families.

He observed that:

• employees were visibly proud;
• their parents and relatives understood the significance of a public-company milestone in a way they might not understand years of technical development;
• the IPO functioned almost like a major coming-of-age or family ceremony.

He recounts the father of one company leader, described as a Chinese immigrant, joking: “I thought it would have happened faster.”

Andrew says the organization consists largely of people who like being “in the trenches” and solving difficult problems.

The IPO therefore gave them a moment of external validation, but after the celebration the internal cultural response was essentially: what is the next problem, and when do we get back to work?

Planet Labs had a different public-market experience.

The host notes that Planet’s share price had moved from roughly $5 to $50 during the preceding 12 months, representing an approximately 10× increase, according to the discussion. This figure should be checked before publication.

Will Marshall says Planet was one of the first public “space stocks.”

He believes investors initially did not understand how profoundly space technology was changing or even how to categorize the company.

Like Andrew, Marshall says the ultimate requirement remains execution of the business.

However, he identifies several real advantages from being public:

• liquidity for early employees;
• liquidity for early investors;
• additional capital for the company;
• greater institutional credibility.

For Planet in particular, credibility matters because it sells to:

• major agricultural customers;
• governments;
• civil-government organizations;
• defense customers;
• intelligence customers.

These organizations may depend on Planet’s data for important functions, so they care whether the provider will still exist years into the future.

Marshall argues that becoming public acts as a maturing and legitimizing event. Customers can see that the company:

• is established;
• has access to capital;
• is less likely to disappear;
• has a visible financial position.

He says some countries are highly dependent on information Planet provides and therefore particularly concerned about Planet’s long-term durability.

Marshall acknowledges that being public means everyone can see the share price every day.

He says Planet does not manage the company around day-to-day stock fluctuations.

Its focus, he says, is building long-term shareholder value.

He believes the public market has gradually begun to understand the scale and importance of space.

He stresses that space infrastructure is already embedded in ordinary life. Examples include:

• satellite communications;
• GPS;
• satellite-derived information incorporated into everyday services.

Consumers may not consciously notice this dependence, but Marshall says the sector is booming.

When asked whether Planet should increasingly be understood as a military or operational-technology company rather than simply a provider of satellite maps, Marshall says the picture is more nuanced.

Planet operates an Earth-imaging satellite fleet.

Marshall claims:

• Planet has the largest Earth-imaging fleet;
• it consists of approximately 200 satellites;
• those satellites image the entire Earth every day.

His analogy is the satellite layer in Google Maps, except Planet’s imagery represents approximately today, instead of an image potentially several years old.

Planet also retains prior daily imagery, creating what Marshall describes as a time series of activity across the Earth.

Potential applications include:

• agriculture;
• energy;
• civil government;
• flood monitoring;
• wildfire monitoring;
• security and defense.

The value is therefore not merely the latest photograph but the ability to compare daily change over time.

Marshall says AI is reducing the barrier to extracting value from satellite imagery.

His reasoning is that satellite data has historically required specialist expertise. AI can make analysis accessible to a much broader set of users.

At the same time, he emphasizes a limitation: AI is only as useful as the data on which it is trained or to which it has access.

This becomes important later in his argument that current language models know a great deal about internet text but comparatively little about changing physical conditions in the real world.

A host asks what percentage of Planet’s revenue/customer base is military or security-related.

Marshall does not provide a numerical percentage in the available transcript.

Instead, he says security was part of Planet’s intended business from the beginning, but acknowledges that it has become a larger fraction than the company originally expected.

He attributes that expansion to current geopolitical conditions.

He argues that Planet’s data can allow governments and security organizations to:

• identify emerging threats;
• see developments “around the corner”;
• receive weeks or months of advance warning;
• potentially take actions that prevent conflict.

Marshall presents this as beneficial to the world because better information may make conflict prevention more likely.

Asked whether he is reluctant for Planet to be perceived as a military company, Marshall says not particularly, but rejects the idea that the company should be understood only in that category.

He emphasizes the broader customer base, including:

• farmers;
• energy companies;
• civil governments;
• NASA;
• security and defense users.

Planet therefore views itself as a broad Earth-data company rather than solely a defense contractor.

Marshall identifies two major technological changes.

He claims rocket-launch costs have fallen approximately four- to fivefold over the past decade.

He says an even less appreciated change is the miniaturization of satellites.

His comparison is that functionality once requiring something resembling:

• a $1 billion satellite
• weighing roughly 20 tons

can increasingly be achieved with spacecraft weighing:

• a few kilograms;
• tens of kilograms; or
• hundreds of kilograms.

He says modern smaller satellites can perform as much work, and potentially more.

Marshall compares this to the transition from mainframe computers to desktop computers.

The important implication of that analogy is not merely lower cost. Mainframe-to-PC miniaturization created entirely new applications because many more organizations and individuals could own computing capability.

He argues that satellite miniaturization is producing a similar explosion of space applications.

Lower launch cost and smaller spacecraft reinforce each other.

The moderator describes Planet and Cerebras as sitting at the beginning of enormous secular technology trends.

The space trend is framed as rebuilding in the sky the data-processing infrastructure previously built on Earth.

The AI-silicon trend is framed as a “rebirth of silicon” in which computing progress may increasingly depend on new architectures and time-to-solution rather than simply traditional transistor-density improvements associated with Moore’s Law.

Cerebras is used as an example of a company that abandoned conventional chip-design assumptions and pursued a radically larger die.

The conversation then asks Marshall to explain whether space-based data centers are genuinely viable.

Marshall first distinguishes the near-term opportunity from the longer-term space-compute thesis.

In the near term:

• satellites are collecting vastly more data about Earth;
• systems such as SpaceX/Starlink and OneWeb are transporting vastly more data;
• capabilities previously available mainly to governments are becoming broadly accessible.

Marshall estimates the potential Earth-observation market at approximately $75–100 billion.

He expects AI layered over Earth imagery to unlock applications across:

• agriculture;
• energy;
• civil government;
• permitting;
• other operational processes.

His thesis is that combining large language/model-based interfaces with Earth imagery will make many existing workflows more efficient.

Marshall says Planet worked with Google approximately eight or nine years earlier on a study comparing:

• ground-based data-center economics;
• the cost of placing computing infrastructure in space.

According to Marshall, the study concluded that orbital data centers become simply cheaper than terrestrial data centers once launch cost falls to approximately $200–$300 per kilogram.

He says current launch cost at the time of the discussion is around or slightly above $1,000 per kilogram.

He also claims launch cost has fallen approximately 10× over the previous ten years.

Marshall predicts that, particularly with Starship, launch prices could reach the critical $200–$300/kg range in roughly two to three years.

He jokes that Elon Musk might predict it for “next week,” but gives his own more conservative expectation of a couple of years.

All figures and timelines are Marshall’s estimates and require external verification.

Marshall says the economics are driven heavily by energy.

Data centers are fundamentally a power problem.

On Earth, solar power can provide inexpensive electricity, but solar generation is intermittent.

To operate continuous computing capacity, terrestrial systems therefore need some combination of:

• batteries;
• gas;
• nuclear;
• other backup generation.

Those additions materially increase the effective cost of reliable power.

In space, Marshall proposes placing solar panels in a sun-synchronous dawn/dusk orbit, where they could remain exposed to sunlight nearly continuously.

He claims this could produce approximately five times more energy per solar panel than a comparable terrestrial panel while avoiding the need for conventional battery backup.

Under his simplified model, an orbital computing system primarily needs:

• solar panels;
• compute chips;
• RF communications links for moving data to and from Earth.

Marshall argues that the key question therefore becomes whether launching the required solar panels and computing hardware is cheaper than building the equivalent reliable infrastructure on Earth.

Marshall says Planet is already participating in early experiments.

He states that:

• Planet has already launched some Nvidia GPUs into space;
• Planet is partnering with Google to launch some of Google’s TPUs;
• an early TPU test is planned or underway.

He acknowledges that substantial technical work remains.

Despite this, Marshall makes a very aggressive prediction: within ten years, most compute will be put in space.

He says the resulting market would involve trillions of dollars and could become larger than any existing space business.

He explicitly links this belief to Planet’s decision to enter the area.

Andrew agrees that space-based compute is important and likely to happen eventually, but he is less confident about Marshall’s timeline.

His main objection is cluster communication.

Modern AI computing does not merely require putting isolated GPUs somewhere. It requires extremely large clusters whose processors must communicate efficiently.

Andrew notes that the industry is still not exceptionally good at solving this problem on Earth, making it even more difficult in orbit.

He nevertheless calls orbital compute an extraordinarily important problem that deserves investment and experimentation.

His uncertainty concerns timing, not the eventual possibility.

Andrew asks whether space compute might be one of those technologies where the last 10% of the engineering problem consumes 80% of the time.

He compares it with self-driving vehicles:

• the industry appeared close to solving autonomous driving;
• the remaining final portion proved enormously difficult;
• completing that final portion required roughly another decade of work.

Andrew therefore warns against treating early demonstrations as evidence that industrial-scale orbital AI clusters are only a few years away.

Despite his skepticism, Andrew says cheaper launch is essential.

Lower launch cost makes repeated experimentation economically possible.

Once experiments become affordable, companies can:

• try architectures;
• discover failures;
• revise designs;
• repeatedly learn from mistakes.

Without sufficiently cheap launches, the feedback loop required to solve difficult engineering problems remains too expensive.

So Marshall and Andrew disagree primarily over how quickly orbital compute becomes practical, not over whether falling launch prices improve the odds that it eventually happens.

The discussion turns to Cerebras and AI chip architecture.

Andrew says AI changed computing by allowing computers to work effectively on classes of information they historically handled poorly.

For most of computing history, he argues:

• computers were excellent at numerical transformation;
• computers could store images but had limited ability to understand them;
• computers could store language but had limited ability to understand or generate it.

Around 2015–2016, AI began changing this.

Computers could increasingly:

• extract information from images;
• understand aspects of images;
• generate language;
• understand language rather than merely store and retrieve it.

This opened huge previously inaccessible areas of knowledge to computation.

At exactly the same time, humanity was producing and collecting vastly larger quantities of images — including both terrestrial and satellite imagery.

Andrew therefore views the AI-compute boom partly as an expansion of what computing is capable of attacking.

This expansion underpins Nvidia’s growth and the wider surge in demand for AI processing hardware.

Andrew says computer architects recognized AI as a possible market-reset event because major new workloads have repeatedly redistributed market share.

He gives several historical analogies.

The rise of computer graphics created demand for dedicated GPUs and enabled Nvidia’s emergence.

When mobile/phone computing became dominant, incumbent leaders such as Intel and AMD did not automatically capture the market despite their manufacturing capability and engineering resources.

Instead, much of the architecture shifted toward ARM.

Andrew also cites a late-1990s networking transition in which older companies lost ground and newer companies such as Cisco, Juniper, and “Arist” [transcript unclear] emerged.

His lesson is that new workloads can break existing competitive structures.

When Cerebras saw AI becoming an enormous consumer of compute, the company believed another such industry reordering was possible.

Andrew says Cerebras made two major bets.

General-purpose compute would not be enough; AI would justify processors designed specifically around the workload.

His reasoning is that if a challenger wants to be 20 times better than an incumbent, merely imitating the incumbent’s architecture is unlikely to succeed.

Nvidia has already optimized GPUs and captured much of the obvious architectural “low-hanging fruit.”

Andrew therefore says that if a company simply builds another GPU, Cerebras’ view was that its odds of outperforming Nvidia were approximately zero.

To obtain a step-change in performance, Cerebras concluded it needed a fundamentally different design.

Andrew identifies a central AI-computing bottleneck as moving data between memory and computation.

Cerebras attacked the problem by:

• building a dramatically larger chip;
• locating memory very close to compute;
• using a different, faster type of memory.

He describes Cerebras’ processor as roughly the size of a dinner plate, compared with conventional chips approximately the size of a postage stamp.

The larger physical design makes it possible, according to Andrew, to use memory and compute in a way that substantially increases effective speed.

He presents this architecture as a deliberate rejection of normal chip-design assumptions rather than an incremental GPU improvement.

Andrew says that when OpenAI uses Cerebras, Cerebras is approximately 15–18 times faster than a GPU.

This is a company performance claim and must be independently checked, including:

• which Cerebras system;
• which GPU;
• which model;
• which workload;
• whether the comparison concerns tokens per second, latency, throughput, or another measure.

Andrew argues that greater speed improves AI in several ways:

• answers arrive faster;
• interaction feels more natural;
• users are more willing to stay engaged;
• faster response allows AI to tackle harder problems without intolerable waiting periods.

Andrew compares AI latency with previous internet transitions.

He asks:

• How large is the market for slow search? His answer: effectively zero.
• How large is the market for dial-up internet? Again: effectively zero.
• How long will a user wait for a website before leaving? He suggests roughly three to five seconds.

His conclusion is that users ultimately will not wait for AI either.

AI therefore needs to be delivered increasingly in real time.

He says that requirement for real-time interaction is what Cerebras identified and built toward.

The discussion returns to public markets and LP liquidity.

Marshall says Planet went public in 2021, via what the transcript renders as a “spa” [transcript unclear; verify intended transaction structure].

A moderator states that Planet entered the public markets at an approximately $2 billion valuation.

When asked to name early investors, the transcript records garbled references including:

• “Drake Professor Jverson” [transcript unclear];
• Capricorn;
• Peter “Te’s” Founders Fund [transcript unclear];
• Yuri Milner’s DST.

All names should be checked against primary records before publication.

The host makes an important public-market observation.

Four years after Planet’s listing, he argues that approximately 90% of the company’s value creation occurred during years three and four, rather than immediately at listing.

Marshall says most early investors remained shareholders and therefore captured much of the later upside.

He specifically says:

• Google had not sold a share at the time of the discussion;
• Google was Planet’s largest single investor;
• Capricorn had held its shares until very recently.

Marshall says those investors benefited from remaining invested and indicates, with acknowledged self-interest, that he believes they should continue holding.

The panel uses Planet to challenge a common LP instinct.

Some limited partners want venture funds to distribute shares quickly once a portfolio company becomes public.

A panelist recounts a separate investment example:

• the investment was made pre-IPO at approximately a $1 billion valuation;
• the fund distributed its shares when the company was valued at approximately $3–4 billion;
• within the following 24 months, the company reportedly reached approximately $50 billion;
• LPs later asked why the fund had not retained the shares;
• the fund’s response was that those same LPs had been pressuring it to distribute.

The example is used to argue that immediate liquidity can conflict with long-term return maximization.

The moderator says Cerebras and its banks created an innovative lockup structure described informally as a “dribble lockup.”

According to the discussion:

• existing investors remained subject to lockup after the IPO;
• shares could begin coming out over approximately six months;
• releases depended on various performance hurdles;
• this staggered mechanism could reduce the risk of all early holders becoming liquid simultaneously.

The speaker predicts that SpaceX could eventually use a similar type of mechanism.

The transcript indicates Altimeter was among the locked-up Cerebras investors.

The details of the structure, terminology, performance hurdles and release schedule should all be checked against formal securities filings before publication.

Asked whether venture investors should distribute their Cerebras holdings immediately after lockup, Andrew argues that historically more money is made after IPOs than before them.

He claims studies show that the post-IPO opportunity is larger both:

• in percentage terms; and
• in absolute dollars.

His reasoning is partly about capital capacity.

Private venture companies often can accept only modest amounts of investment relative to later public companies.

Even if the percentage return in an early round is attractive, the investor may have been able to deploy only a relatively small amount.

Once a successful company is public, there can be much more equity available and therefore much more capital that can participate in subsequent appreciation.

Andrew’s conclusion is that the absolute wealth-creation opportunity can therefore be substantially larger after the IPO.

This is presented as Andrew’s investment-market thesis, not independently verified fact.

Marshall or another panelist then applies the argument to SpaceX.

Historically, major technology companies often went public at valuations of only a few billion dollars.

They subsequently created enormous value in public markets.

SpaceX, by contrast, is discussed as potentially entering public markets at a valuation measured in trillions.

The speaker observes that there are many orders of magnitude between a few billion and a few trillion.

For SpaceX to reproduce the same post-IPO multiples enjoyed by earlier technology giants, it would theoretically have to reach valuations in the quadrillions.

The speaker jokes that Elon Musk may have ambitions of that scale, but the point is serious: a huge private-market valuation can transfer much of the early growth from public investors to private investors before the listing ever occurs.

The panel contrasts giant private companies with Planet Labs.

Planet is presented as an example of venture-like return occurring in public markets.

The host notes Planet’s approximately 10× public-market move.

He argues that companies going public earlier allow ordinary public-market investors to participate in the high-growth phase.

He makes the same counterfactual argument about Cerebras:

• if Andrew had been able to take the company public roughly 18 months earlier;
• it might have listed at approximately $10 billion rather than $50 billion, according to the speaker;
• the subsequent roughly 5× increase would then have accrued to public shareholders rather than only private investors.

Andrew responds with the remark: “Way better to be lucky than good.”

These valuations and counterfactuals are speakers’ claims and estimates.

One speaker rejects the idea that companies such as:

• Anthropic;
• OpenAI;
• SpaceX

necessarily represent the permanent new model in which technology businesses remain private until enormous valuations.

He believes the pendulum may be moving back toward earlier IPOs.

He says portfolio companies are once again considering listings at valuations such as:

• $1 billion;
• $3 billion;
• $5 billion.

The discussion contrasts this with roughly a decade in which “Andrees” [transcript unclear] and others strongly promoted the idea of remaining private for much longer.

The speaker now sees founders becoming more interested in following a Planet-like path: become public earlier, compete in the public markets, and grow under public scrutiny.

A panelist argues that becoming public earlier does more than democratize access to upside.

It also increases competitive and operating discipline.

Both Cerebras and Planet have faced visible competitors and continuous pressure.

Public companies additionally face:

• investor scrutiny;
• quarterly and financial scrutiny;
• visible market judgment;
• pressure to deliver.

The speaker believes this pressure sharpens organizational focus.

He describes the mechanism with the repeated metaphor that steel or iron sharpens steel/iron.

His view is that innovation can improve when a company must perform visibly and is subjected to external discipline.

The broader argument is therefore twofold:

1. earlier IPOs allow more people to participate economically;
2. public accountability may improve company execution.

Another panelist explicitly agrees.

The moderator closes by praising both Andrew and Marshall as examples of entrepreneurship.

He refers to having known or followed their companies from very early stages and notes that different competitive and strategic paths later converged around major technology trends.

The panel expects the following several years to be unusually active or, in the moderator’s wording, “really spicy.”

This leads Marshall to offer a final technological thesis.

Marshall argues that the present generation of large language models is extraordinarily powerful but fundamentally constrained by its source material.

Much of their capability comes from ingesting the text of the internet.

That gives the systems extensive linguistic and informational knowledge.

But, in Marshall’s characterization, they remain “blind” to the real world.

They do not inherently know the current condition of:

• a particular farm field;
• a developing flood;
• a fire;
• a security situation;
• some physical event “around the corner.”

Satellite systems can provide the missing real-world data.

Once AI models have access to changing physical-world information, Marshall believes they can answer real-world operational questions, not merely questions about recorded internet text.

He expects this to create an enormous number of new applications.

Marshall proposes new language for this convergence.

Instead of only large language models, he describes the possibility of “large Earth models.”

Instead of thinking only in terms of artificial intelligence, he proposes “planetary intelligence.”

His envisioned system contains:

• planetary sensing systems in space;
• continuous observation of the physical Earth;
• AI capable of understanding those observations;
• eventually, planetary compute systems in space as well.

Marshall acknowledges that reasonable people can disagree about the exact timeline for orbital computation.

His conviction is about the direction rather than the precise schedule.

He believes the combination of:

• AI;
• real-world Earth data;
• ubiquitous sensing;
• cheaper launch;
• orbital computing

will produce an enormous new economy.

He closes by describing the next few years as an especially exciting period for the merger of AI and space.

Exact quote: “Space and AI are really um a match made in heaven. They're getting married.”

Context: Opening description of the convergence between satellite infrastructure, Earth data and artificial intelligence.

Topic: AI + space convergence.

Exact quote: “Just like Google figured out how to index the internet and make it searchable, we are indexing the earth and making it searchable.”

Context: Marshall’s description of Planet Labs’ strategic role.

Topic: Planet Labs / searchable Earth data.

Exact quote: “I believe that the wave is the biggest wave in the history of technology.”

Context: Opening statement about the current technological cycle.

Topic: Scale of the AI/technology boom.

Exact quote: “It’s really difficult to overestimate the amount of garbage that's involved in going public.”

Context: Description of the administrative burden of an IPO.

Topic: IPO process.

Exact quote: “Not a damn thing changes in the important parts of your business.”

Context: His argument that an IPO does not solve operating problems.

Topic: Public markets versus company execution.

Exact quote: “The answer is by getting it wrong for a decade.”

Context: Asked how Cerebras got its IPO timing right.

Topic: Persistence and market timing.

Exact quote: “I thought it would have happened faster.”

Context: A Chinese immigrant father reacting to the company finally going public after a long journey.

Topic: Employee/family meaning of an IPO.

Exact quote: “They want to know you're going to be around.”

Context: Explanation of why large governments and institutional customers value public-company credibility.

Topic: Institutional trust.

Exact quote: “AI is only as good as the data it's trained upon.”

Context: Discussion of applying AI to Earth-observation data.

Topic: AI data limitations.

Exact quote: “What this does is enable them to see threats around the corner.”

Context: Security and geopolitical use of Planet imagery.

Topic: Satellite intelligence / conflict warning.

Exact quote: “We are rebuilding the data processing infrastructure that has existed on the earth in the sky.”

Context: Framing of the orbital-compute opportunity.

Topic: Space-based infrastructure.

Exact quote: “It’s a power game.”

Context: Explanation of why energy economics are central to orbital data centers.

Topic: Data-center power economics.

Exact quote: “Within 10 years most compute will be putting in space.”

Context: Marshall’s strongest orbital-compute prediction.

Topic: Future of data centers.
Note: Grammar preserved from transcript.

Exact quote: “The last 10% is 80% of the time.”

Context: Warning that orbital compute may encounter a difficult final engineering phase.

Topic: Technology-development timelines.

Exact quote: “If you build a GPU, the odds that you're better than Nvidia in our view are approximately zero.”

Context: Explanation of why Cerebras chose a radically different architecture.

Topic: AI-chip strategy.

Exact quote: “The hard part is moving data from memory to compute.”

Context: Cerebras’ identification of the central AI-hardware bottleneck.

Topic: AI semiconductor architecture.

Exact quote: “So when OpenAI uses us, we're 15 or 18 times faster than a GPU.”

Context: Performance claim for Cerebras hardware.

Topic: Cerebras speed / OpenAI.

Exact quote: “You will not wait for AI.”

Context: Argument that AI latency will become as unacceptable as slow web search or dial-up connections.

Topic: Real-time AI.

Exact quote: “Historically more money's made after IPO than before.”

Context: Argument against automatically distributing venture holdings immediately after public listing.

Topic: Public-market value creation.

Exact quote: “Way better to be lucky than good.”

Context: Reaction to the counterfactual that an earlier Cerebras IPO could have allowed the public market to capture more of its rise.

Topic: IPO timing / market luck.

Exact quote: “They don't know about the real world.”

Context: Marshall’s criticism of models primarily trained on internet text.

Topic: Physical-world AI.

Exact quote: “Instead of having large language models, large earth models.”

Context: Proposed model for combining AI with planetary observation.

Topic: Large Earth models.

Exact quote: “Instead of AI, planetary intelligence.”

Context: Marshall’s name for a future system combining sensing, AI and space computing.

Topic: Planetary intelligence.