SpaceX Launches Largest Ever IPO | OpenAI Files to Go Public | Uber Cuts 23% of HR
Summary
- SpaceX’s $75 billion IPO at $135 a share—roughly a $1.8 trillion valuation—replaces price discovery with Elon’s conviction. With the book only 2X covered versus the traditional 8–10X target, Jason Lemkin expects “nominally…a dud,” while Rory O’Driscoll assigns equal odds to a fall, flat trade, or retail-driven rise on day one. Over 12 months, Rory expects valuation to reassert itself and thinks even $1 trillion would remain an extraordinary outcome.
- A flat SpaceX debut would matter less to SpaceX than to the capital-hungry companies following it through the public-market door. Jason thinks it could restrain OpenAI’s valuation or fundraising ambitions, while LPs receiving historic distributions may demand routine 7–8X fund returns and conclude that “little $5 to $8 billion IPOs” no longer move the needle. Rory’s pushback: trillion-dollar results cannot become the operating assumption; fund size determines how large an exit must be.
- Persistent AI, not another browser tab, is the product destination—and Dreaming V3’s memory upgrade is part experience improvement, part token economics. Jason expects today’s non-persistent AI to feel “almost archaic” within two years; Rory says memory should produce better answers while avoiding repeated transmission of the full context. Apple’s Google-powered AI is therefore pragmatic rather than surrender: “What matters for us, Apple, is delivering an amazing experience,” using handset, calendar, and personal context that OpenAI must compete against.
- AI’s clearest labor signal is not Uber cutting 23% of HR but startups designing for radically higher revenue per employee. Lovable was discussed at $500 million ARR alongside a roughly 172-person example, while Rory predicts startups will become roughly half their former size and target at least $1 million in revenue per employee. Rory’s qualification is load-bearing: businesses spending 50–70% of revenue on model “intelligence” cannot also carry legacy labor ratios, and enterprise selling will still require more people than PLG.
- Jason viewed Elon’s AI strategy as a rapid conversion of expensive capacity into a vertically connected compute-and-application stack. After committing an estimated $20–30 billion ahead of revenue, building Colossus and Colossus 2, and initially missing with the model, Elon reportedly secured about $2 billion monthly in Anthropic and Google compute revenue and added Cursor to consume capacity. “Did he turn a loss into a win in the space of three months.”
- The founder revolt against VCs reflects genuine personal injury, but the panel rejected the idea that ordinary fundraising rejection deserves a permanent grievance. Jason’s blunt framing was “get over it because it’s sales,” although he distinguished rejection from being fired; Rory countered that founders are selling themselves, not merely a product, while VCs structurally reject 99 of every 100 opportunities. The Cloudflare–Vinod Khosla dispute illustrated the tension between direct team assessment and needless damage.
- Ramp, Revolut, Suno, and Bending Spoons are all being valued on continued execution while capital remains brave. Ramp’s $44 billion round works if its growth persists; Revolut’s $115 billion price depends on growth-adjusted financial-services economics; Jason finds Suno useful but “fragile”; and Bending Spoons turns acquired customer inertia into cash through cost cuts and price increases. Rory’s cycle summary: “There’s always money when people aren’t afraid.”
- Databricks can remain private because its capital requirement is manageable, but the foundation-model market cannot defer its structural reckoning. Microsoft’s web-blind new models raised doubts about whether even incumbents can catch Anthropic’s pace, while Rory asked whether the market becomes an “oligopoly” or retains four or five credible suppliers in two years. Non-Chinese US open source matters because model concentration determines pricing power for the entire AI stack.
Deep dive
1. SpaceX’s fixed-price IPO abandons the usual pop machinery
At Tuesday’s recording, Elon had already selected $135 per share, valuing SpaceX near $1.8 trillion while seeking $75 billion. Rory’s description: “We’re not doing price discovery. I’m telling you the answer,” leaving investors only to decide quantity.
Jason considered a book only 2X covered weak against the traditional 8–10X target, although Rory noted that achieving 10X demand on $75 billion is unusually difficult. Fixed pricing removes the late demand information bankers normally use to engineer a 10–15% opening gain.
Rory therefore saw a “non-trivial chance” of a downside opening, without claiming his illustrative 30% was knowable. Jason’s conditional call—especially if 30% goes to retail—was a subdued first week followed by an “inexorable rise” whenever launches, satellites, or revenue-linked announcements renew enthusiasm.
2. The medium-term valuation matters more than launch-day optics
Rory refused to let the mechanics obscure the achievement: SpaceX is “the iconic company of its generation,” an extraordinary technical business and “an only in America moment.” His valuation skepticism coexists with admiration for the capital, risk tolerance, market depth, and execution required.
Harry demanded closing prices; Rory assigned one-third odds each to down, flat, or up on day one because the mechanism supplies little information. His firmer 12-month view was bearish: at roughly 70 times forward sales, he doubts the company retains $1.7 trillion, while stressing that $1 trillion would still be “a huge win.”
Jason noted that Facebook and Google debuted “with a whimper” without damaging their eventual outcomes. SpaceX will still create generational wealth and liquidity; the more exposed follower may be OpenAI, whose aggressive capital requirement could meet a cooler valuation or smaller raise after a weak debut.
Jason’s LP heard 7–8X fund expectations and questioned whether $5–8 billion IPOs still work mathematically. Rory rejected extrapolating a singular SpaceX result, contrasting its 2008 investment vintage with a likely Mercer seed check three years earlier: rare trillion-dollar outcomes happen, but fund size—not a new universal base rate—dictates the exit required.
3. SpaceX liquidity will inspire LPs without changing venture arithmetic
Harry asked whether SpaceX distributions would drive more direct investing and larger fund commitments. Rory agreed recipients will understandably chase the next one, but called that an extrapolation from “the best venture capital deal ever in terms of absolute return.”
Ontario Teachers—not “Ohio Teachers,” as Harry first said—was the highlighted institutional winner. Rory also recalled a Journal report involving Washington University and roughly 10–15% of its endowment, as evidence of the concentration and magnitude of the payout.
Rory’s governing constraint came from the old “venture arrogance index”: a larger fund requires a larger company to return it. An $8 billion result can make smaller funds “perfectly bloody happy”; it only becomes inadequate when the vehicle itself approaches $10 billion.
4. OpenAI is filing for optionality as AI becomes persistent
Jason’s question was why OpenAI would file while hedging the timing. Rory read it as overdue expectation management: signal the intention, avoid committing to November, and prevent any ordinary delay from becoming a stream of “WTF is going on?” stories.
Behind the public caveat, Rory expects finance and legal to hear “Get this puppy done as quickly as possible so we have maximum optionality.” SpaceX’s unusually rapid SEC process reinforced his sense that multiple capital-intensive companies are “gunning for the door” while markets remain risk-on.
Harry framed Dreaming V3 as OpenAI’s biggest memory-architecture upgrade since launch. Jason believes AI living primarily inside browsers already feels dated and expects non-persistent interaction to look “almost archaic” within two years, even as finite IT budgets force discipline around token consumption.
Rory placed memory inside the broader model “harness”: persistent context should improve responses and reduce costs because the system need not repeatedly pass everything through the frontier model. After researching 58 episodes, he joked, OpenAI should know he is probably returning for more 20VC research.
5. Apple’s distribution may matter more than owning the model
Jason called Apple’s Google arrangement a form of giving up; Rory disagreed. Apple may pay Google about $1 billion for the model while receiving $20 billion for default search, making the AI expenditure a small offset rather than strategic capitulation.
Rory conceded Apple “screwed up” by lacking its own model, but argued that handset control supplies richer context—identity, calendar, history, and intent—than a standalone subscription. Fixing Siri with an external model is progress if the resulting experience keeps customers buying devices.
Borrowing Ben Thompson’s framing, Rory distinguished enterprise productivity from consumer leisure: “Consumers don’t wanna work.” Anthropic’s enterprise bet fits automation and efficiency; OpenAI’s consumer position must compete with Apple and Google on delightful experiences for people who often want relaxation, not complex research.
6. Uber’s HR cut is a noisy AI signal; robotaxis are the larger bet
Harry highlighted Uber cutting 23% of HR, restoring a three-day office mandate, and denying an AI connection despite 95% of engineers using it. Jason separated recruiting—habitually cut whenever growth stumbles—from HR functions that he thinks AI could manage more comprehensively and sometimes less prejudicially.
Jason’s proposed “AI VP of HR” could process every work product, complaint, and pattern rather than relying on a partial human view; it might even conclude “maybe it really is your idiot boss.” He explicitly stopped short of advocating removal of every human or attributing Uber’s entire cut to automation.
Rory doubted non-engineering adoption was strong enough to generate 23% savings by itself. He treated the layoff as one data point in the disputed range—5%, 10%, or Dario’s 50%—between automating knowledge-work tasks and eliminating complete jobs, with ordinary overstaffing likely mixed in.
The more material Uber signal was robotaxi experimentation in Madrid with a partner Rory thought might be WeRide. Autonomous driving has advanced far more slowly than the expected domino effect; that gives Uber time to turn robotaxis from an existential threat into fleet supply coordinated through the app consumers already use.
7. Revolut’s $115 billion value indicts incumbent banking
Harry offered Revolut’s $115 billion valuation as Europe’s rebuttal. Rory embraced the company but credited its opening to “fat, dumb, and happy” European banks extracting excessive margins, especially through historical FX and cross-border charges that Revolut could undercut.
The same framework explained New Bank’s opportunity against inefficient Brazilian incumbents and Chime’s roughly $5 billion value in a more efficient US market. Fintech outcomes, in Rory’s view, scale with how “egregiously priced” the legacy providers were.
Rory’s longer-run caveat was systemic: a bank that becomes largest by market capitalization without doing much long-term lending may be commercially excellent but does not perform banking’s core economic function—recycling savings into credit. He bracketed that concern because Revolut’s current execution remains exceptional.
8. Fundraising pain is personal, but rejection is the venture default
Jason acknowledged that perceived slights “really burned” during his founder years, including investors who repeatedly used him to diligence competitors. His mature advice is to take that meeting anyway and conduct “reverse intel,” because founders retain grudges while VCs simply look for another bus after missing a deal.
His harder conclusion was “get over it because it’s sales.” A promised customer can ignore 28 emails and 87 texts just as an investor can abandon a stock purchase; the grievance he treats differently is being fired, citing former Uber executives’ anger toward Benchmark as understandable.
Rory’s pushback—worth keeping—is that the founder is selling the self, not a Ford car, so rejection lands personally. Yet a venture firm rejects roughly 99 of 100 reviewed companies, just as a prudent bank declines five of six borrowers; high satisfaction is structurally difficult when “no” is the default product.
Harry argued the best revenge is forgetting rejectors existed, particularly when they return later. Rory conceded thick skin accumulates, though he still vividly remembers three LP rejections within one hour after timing his first independent fundraise for the November 2008 financial crisis.
9. The Cloudflare dispute shows how direct advice becomes lasting damage
In the circulated Cloudflare story, Jason understood Vinod Khosla to have suggested removing Michelle and the CTO and reallocating shares—not stealing them. Jason would not have made that recommendation, especially during a pitch, but recognized the underlying problem investors sometimes perceive: an uneven founding team.
Rory emphasized the revealed outcome: whatever Cloudflare possessed “shouldn’t have been touched one little bit.” He also noted that Khosla denied the exchange happened, and that a highly successful, exceptionally direct investor may choose one blunt conversation where three tactful meetings would have caused less damage.
Khosla appeared on the Midas List for Juniper and, three decades later, for OpenAI; overall ability and an offensive individual meeting can coexist. Rory’s standard is neither denial nor permanent condemnation: acknowledge inevitable breakage across hundreds of annual rejections, apologize when wrong, and move on.
10. Lovable turns revenue per employee into a strategic choice
Lovable was discussed at $500 million ARR alongside a roughly 172-person example; Cursor reached $4 billion and targeted $6 billion by year-end. Jason sees the unresolved question as whether AI startups remain lean through $100 million, $500 million, and $1 billion—or eventually “get fat again” with layered organizations.
He rejected the idea that these are trivial single-product companies: coding platforms ship databases, hosting, management, SEO, and relentless features in an intensely competitive category. That productivity made him “kind of contemptuous of startups that need to be fat”: executives demanding another 50–100 people or $10–40 million should often leave.
Rory’s qualification was economic rather than cultural. A company spending 50–70% of revenue on Anthropic or OpenAI intelligence cannot spend the same share on employees; tokens and model intelligence change the labor mix and let a small group capture unusually high leverage, compensation, and revenue per head.
PLG can stay lean because “people are either making stuff or selling stuff”; Rory said enterprise distribution generally requires a larger sales force. He and Jason debated whether new companies will recreate Oracle-scale sales organizations: Replit is hiring 250 salespeople while Lovable is not, and Rory said some founders may trade marginal revenue for two-to-five-times efficiency. Jason agreed that average efficiency will improve but rejected the idea that enterprise sales can be handled by 147 people.
11. AI cost structure makes old and new revenue-per-head ratios incomparable
Harry contrasted more than $3 million ARR per employee with Salesforce near $350,000. Jason then noted that Salesforce spends little on tokens—roughly 1% of revenue in his illustration—while a business such as Replit might generate $2.3 million per head but spend roughly 70% of revenue on them.
Rory said that founders today want at least $1 million in revenue per employee, aspire to $2 million, and want small teams of exceptional colleagues. His directional prediction is that startups, including B2B companies, will operate at roughly half their historical headcount for comparable revenue.
Jason accepted the logic categorically: if AI eventually generates $1 trillion of revenue by augmenting people, the corresponding efficiency must appear as fewer humans per unit of work. “If the people who sell AI can’t be efficient with AI, then what chance is there for the rest of them?”
12. Jason says Elon converted AI capacity into a compute-and-Cursor stack
Jason called the Cursor acquisition clever “on every dimension.” Over roughly 24 months, Elon moved from a standing start through Colossus and Colossus 2, suffered a model failure, yet committed an estimated $20–30 billion before revenue because he believed AI was the trend to back.
That bet reportedly left him with gigawatts of capacity at precisely the moment competitors needed it. Jason cited approximately $2 billion a month—$24 billion annually—from Anthropic and Google, while Cursor’s targeted $6 billion business could provide additional downstream demand for those servers.
Jason’s distinction remained sharp: this does not prove a winning foundation model; it makes Elon “a better CoreWeave” with exceptionally cheap capital. But the narrative flipped from stranded data centers on January 1 to a large outsourced-compute business and owned application demand by June 9.
13. Risk-on capital rewards growth and punishes small misses
Ramp raised $750 million at $44 billion after tripling, crossing $1 billion ARR, and reaching positive free cash flow. Jason estimated a 30–40X revenue multiple if revenue is near $1.25 billion: sensible if growth persists, indefensible if it falls toward “normalized growth.”
Revolut similarly reported about $4.5 billion in revenue and $1.5 billion in operating income, yet ordinary banks trade nearer 12X earnings than 40–50X. Jason’s answer to “tech or financial-services multiple?” was financial-services economics adjusted upward for exceptional growth.
Suno raised $400 million at $5.4 billion, twice its valuation six months earlier. Jason pays $15–20 monthly despite using it for perhaps three songs; that makes the product impressive but “fragile,” and he cannot yet see the pathway to the implied $20 billion outcome.
Rory located the funding source in psychology: “There’s always money when people aren’t afraid.” Jason warned that 1-to-100 growth narratives still encounter GDP; Rory added human nature and Minsky-style overreach. Broadcom’s $16 billion chip guidance missed the $17.2 billion expectation and helped trigger the market dip.
14. Bending Spoons monetizes inertia rather than pure organic growth
Bending Spoons filed around a $20 billion valuation with roughly $1.3 billion in revenue after acquiring Evernote, Vimeo, WeTransfer, AOL, and Eventbrite. Jason admired the strategy and joked that if AOL becomes the next hot thing, “these guys are fucking geniuses.”
Rory’s closer reading produced a different definition of turnaround: remove low-return acquisition spending, reduce teams, focus features, and raise prices aggressively. For an illustrative property, an 80% price increase might lose 10–20% of users while retaining enough deeply embedded customers to expand revenue and cash flow.
Evernote’s average price reportedly rose from $75 to $250 annually; AOL is the purest inertia asset because “anyone who hasn’t churned from AOL now ain’t churning until they die.” Rory saw the playbook as consumer Vista or Thoma Bravo—highly profitable, but not necessarily worth 15–20X revenue when acquisitions drive much of growth.
Jason was less concerned about the organic label if targets remain affordable and execution repeats across “800 unicorns.” He also liked the founder letter’s distinction: discovering product-market fit contains luck, while operating an acquired product after fit can become a repeatable machine.
15. Databricks can wait, while model-market structure cannot
Databricks chose another round at $165 billion, up from $134 billion earlier in the year. Rory reduced the IPO decision to three needs—capital, acquisition currency, or shareholder liquidity—and argued a software company without foundation-model-scale spending can remain private when terms are cheaper and less burdensome.
He would generally favor going public around $4–5 billion in revenue, but private investors currently offer Databricks a higher growth multiple than Snowflake receives publicly. With SpaceX and two model providers crowding the calendar, waiting for a cleaner year can be rational.
Jason found Microsoft’s newly launched models revealing because they could not search the web. Even if that omission fits some use cases, it revived the stale-knowledge era of early ChatGPT and challenged the assumption that Microsoft, DeepSeek, or open source can automatically keep pace with Anthropic.
Harry argued many open models already sit within striking distance; Rory replied that they are mainly Chinese and some providers may move closed source. The consequential two-year question is whether Reflection AI, Poolside, or another US alternative prevents an OpenAI–Anthropic oligopoly, preserving pricing pressure and the broader ecosystem Nebius says it needs.