Groq’s $20BN NVIDIA Deal | Why Sam Altman Doesn’t Care About Dilution & Invisible Unemployment 2026

Summary

• NVIDIA’s $20 billion Groq deal was framed as strategic insurance on its extraordinary economics, not a revenue-multiple acquisition. Always-on agents shift AI’s center of gravity from one-time training toward constant inference, where Groq offered predictable, low-latency performance. At roughly 3x its last-round valuation but less than 20% of NVIDIA’s stated annual cash generation, the deal removes one of perhaps five or six credible sources of margin pressure: “This was a poker game.”
• Cerebras inherits both the industry’s best valuation comp and a newly strengthened competitor. Groq’s price could reset how corporate buyers and IPO investors value scarce silicon assets, creating “a damn unlocking” inside acquisition committees. Yet Cerebras also lost NVIDIA—the obvious first name on its buyer slide—and now faces the “musical chairs” problem of finding another acquirer willing to pay strategically rather than financially.
• Meta’s $2.5 billion Manus acquisition may have captured the founders’ local maximum even if it was suboptimal for Benchmark’s fund math. Manus had reached $100 million of ARR and a $125 million consumption-inclusive run rate, making the price 25x current ARR after roughly 5x growth in eight months; the team was said to still own about 80%. Harry saw an underpriced asset, but Jason and Rory emphasized competition, likely low gross margins and founder concentration: “Words are words and half a billion dollars is life-changing.”
• Meta and OpenAI are both spending aggressively because winning AI talent matters more than preserving tidy economics. Zuckerberg appears willing to risk hundreds of billions rather than remain irrelevant, while OpenAI reportedly spends 46% of revenue on stock compensation—$1.5 million per employee, 34x comparable pre-IPO companies—yet retains only roughly 60% of researchers. The governing maxim was: “No one ever said to Winston Churchill, ‘Congratulations, you won World War II on budget.’”
• Permanent, personalized AI could justify today’s apparently excessive compute, power and hardware bets. Jason’s Claude searched 14 months of his work and personal context, named itself “Ren,” and became his model for an assistant that accompanies users continuously; Rory accepted the 24/7 workflow thesis while rejecting literal sentience. If knowledge workers run several agents all day, “you need 1,000 times what we have today,” making inference infrastructure the load-bearing investment case.
• Navan trading near 4x ARR suggests the IPO window is barely open for companies without an AI premium. The company was described as growing 27–28%, cash-flow positive and non-GAAP profitable, but its timing, CFO departure, roughly $700 million of debt and $200 million of cash made an IPO the “least-bad option.” Private capital still appears cheaper despite being illiquid, leaving public markets with an uncomfortable question: “You guys just aren’t a compelling product.”
• Cash generation increasingly lets scaled private companies treat an IPO as optional rather than inevitable. Revolut was cited at $9 billion of revenue and $3.5 billion of 2025 profit, with a $75 billion private valuation versus Chime’s fall from $25 billion privately to roughly $6 billion publicly; dividends can provide founder liquidity without selling shares. For Databricks, however, public stock could unlock $10–50 billion acquisitions that remain awkward while private.
• “Invisible unemployment” is the episode’s darkest 2026 call: headline data may lag while entry-level and late-career opportunities come under stress. Companies are holding headcount flat, backfilling with AI and eliminating junior sales and knowledge-work roles, while older executives quietly discover that their 2021 toolkit no longer clears the market. Elite researchers still command millions, but the rest face a harsher barbell—and political backlash becomes plausible when graduates conclude that “these guys have built a future that doesn’t need you.”

Deep dive

1. NVIDIA bought strategic insurance against an inference-first world

• The panel’s premise was that knowledge workers will soon run multiple agents continuously—effectively consuming 48 or 72 hours of inference per day. Training happens once; inference happens on every query, making it “all of the growth” as AI becomes a permanent operating layer.
• Elad located Groq’s advantage narrowly but valuably: predictable, low-latency inference. Tavus used it because conversational digital presences cannot freeze while waiting for responses; that performance matters much more when an assistant is always listening, reasoning and acting.
• NVIDIA was described as charging a small customer base roughly 75% gross margins while producing about $100 billion of annual cash. Paying $20 billion—less than 1% of its market capitalization and under 20% of annual cash generation—to eliminate one of few plausible margin threats therefore looked rational.
• The revenue history made ordinary valuation analysis useless: under $4 million in 2023, roughly $40 million in 2024 and about $175 million more recently. “This was a poker game”: an asset perhaps worth $5 billion independently could be worth multiples more to the company whose franchise it protects.

2. Speed, scarcity and structure explain Groq’s 3x premium

• The reported sequence was extraordinary: Jensen Huang contacted founder Jonathan Ross only weeks before closing and demanded completion before Christmas. Harry’s experience was that a buyer arriving “hot” can clear objections instantly by paying roughly 3x the last-round price.
• Groq’s pedigree mattered. Ross had helped create Google’s TPU before founding the company around 2016–17, and Chamath Palihapitiya’s Social Capital backed that “S-tier leader” through a long semiconductor winter before AI compute became exceptionally valuable.
• Elad inferred that NVIDIA also paid for “transactional compliance.” A conventional acquisition of a vaguely comparable competitor might attract regulatory scrutiny; a rapid license-and-hire-style transaction could secure the technology and team immediately, even if speed required a higher price.

3. Cerebras gained a benchmark and lost its first-choice buyer

• Cerebras, which had reportedly raised about $1 billion near a $5 billion valuation and was considering an IPO, now owns “the world’s best comp.” Bankers can argue that strategically scarce AI silicon deserves much more than a conventional semiconductor multiple.
• Jason added a subtler benefit: a $20 billion precedent resets corporate psychology. Once the industry’s most respected semiconductor operator pays that price, acquisition committees stop laughing at similar proposals and begin asking what winning requires—“a damn unlocking” of budgets and ambition.
• The negative is literal musical chairs. NVIDIA has bought its preferred asset, while Google already has TPUs; Cerebras may now depend on Amazon, Apple, Microsoft or OpenAI deciding that escaping the NVIDIA tax requires an owned silicon strategy.

4. Groq remains a singular semiconductor outcome, not a new playbook

• Rory credited Social Capital for the early rounds, followed later by crossover investors such as D1 and Tiger. Most established venture firms had abandoned chips after brutal outcomes from roughly 2000 onward.
• Elad’s caution was categorical: “I don’t think this means that there are going to be 20 more semiconductor great outcomes.” Groq survived long enough for a singular AI wave to meet exactly the right technical asset; that does not make the category broadly venture-friendly.
• His analogy was Arista Networks: one exceptional networking company emerged after VCs stopped funding “box companies,” but it did not reopen an enduring conveyor belt of similar winners. Elad’s conclusion was to celebrate the outlier without manufacturing a sector rule from it.

5. Manus sold at a credible local maximum

• Harry supplied the transaction specifics missing from public reporting: Meta paid $2.5 billion, or 25x Manus’s $100 million current ARR; consumption lifted its run rate to $125 million, after roughly 5x growth in eight months.
• Benchmark’s bet had been genuinely non-consensus. It funded what initially appeared to be a China-based company, then Manus moved its base to Singapore and severed relevant links, reducing geopolitical and corporate-structure risk before the exit.
• Rory questioned the obvious product fit between a knowledge-worker tool and Facebook’s roughly three billion users. The strategic asset was the team’s demonstrated ability to make complicated AI orchestration work for ordinary, nontechnical users.
• Jason and Rory’s “local maximum” case combined competitive and economic risk: Anthropic, OpenAI, Replit, Lovable and Base44 could converge on similar workflows, while running multiple external models likely constrained gross margins. An epic product was not necessarily an enduringly defensible standalone company.

6. Founder wealth and venture fund returns optimize for different maxima

• The team was said to retain about 80%, putting the founders in line for hundreds of millions each, with potentially no Singapore capital-gains tax if they were Singapore residents. Harry’s practical test was: “Normal people would have taken it at $800 million.”
• Harry’s broader pushback was that young founders might lack investors’ market comparisons. Rory added that, at continued 3x growth, the price could equal only about 8x the following year’s revenue; the board-level case for waiting was therefore not absurd.
• Jason challenged the argument by suggesting a $10 billion secondary: if investors really believed the company was underpriced, they could buy $500 million from each founder. Diversified VCs can recommend compounding; concentrated founders bear the existential risk. “Words are words and half a billion dollars is life-changing.”
• The fund math reinforced the misalignment. Jason called a roughly one-third-of-fund return unexciting, while Rory valued putting a 4x win on the board within six to eight months—especially in a fund he described as roughly 15x. Neither outcome determines what the founders should choose.

7. Forcing an unwilling founder to hold is worse than selling early

• Jason’s operating rule was that founders control exit decisions roughly 90% of the time, and VC intervention in the remaining cases is usually a mistake. A board can surface risks and ask about “nagging worries” the CEO has suppressed, but it cannot diversify the founder’s life.
• Rory’s counterexample was Cruise: autonomy may eventually create a $100–200 billion company, yet selling for roughly $1 billion in 2016 could still have been exquisitely timed because the sector produced no comparable exit for years.
• Keeping a reluctant CEO becomes toxic if the company later deteriorates. The founders might be naive—but they might also understand their specific competitive position better than investors armed with broader comps. Rory’s conclusion was that they “may actually be incredibly astute to sell here.”

8. Meta’s AI reset mixes managerial urgency with “spite”

• Harry described Yann LeCun’s FT interview as unusually explosive: LeCun called Alexandr Wang young, naive and inexperienced, while separately suggesting that Llama performance had been presented through selective or incorrect benchmarks during LeCun’s own tenure.
• Rory saw a fundamental employer-employee mismatch. LeCun believes LLMs alone will not reach AGI; Zuckerberg needed someone to ship a model competitive with OpenAI and Anthropic now. Once Llama became a corporate imperative, a leader whose opening position was “this isn’t important” could not remain aligned.
• Jason called this “the era of the spite startup,” placing Anthropic, xAI and LeCun’s new effort in Silicon Valley’s tradition of Fairchild and Intel: talented people leave, take their marbles and try to prove their former institution wrong. Rory’s hedge: “Spite might provide motivation. It doesn’t guarantee outcomes.”
• The structure improved Rory’s view—LeCun as chairman, with Alexandre Lebrun, formerly CEO of Nabla, as CEO—but not his category outlook. The scientists may be top 0.01%; the unanswered question is whether markets can support ten costly research labs that each must also build a business.

9. OpenAI is diluting heavily because talent remains the constraint

• OpenAI reportedly spends 46% of revenue on stock compensation, about $1.5 million per employee and 34x comparable pre-IPO technology companies. Jason’s provocative explanation was that a CEO with no shares has little reason to prioritize dilution over building “the biggest, greatest AI” company.
• Rory thought Sam Altman might be right anyway. With Meta offering key people $20–50 million in liquid stock, preserving a clean cap table is secondary to winning: “No one ever said to Winston Churchill, ‘Congratulations, you won World War II on budget.’”
• Reported compensation also understates employee economics when valuations rise. A grant booked at $10 continues amortizing from that price even if the shares reach $40; economically, someone showing $1.5 million of expense might be receiving $4–5 million.
• The more revealing measure may be annual ownership transferred: Jason said he would not be surprised if some of these companies gave away 8–10% annually, versus 2–3% for slower-growing public companies. Anthropic’s headline valuation rose about 15x from $4 billion to $60 billion, yet Rory’s rough math implied only about 5x per share after financing and employee dilution.

10. Masa’s concentration makes the OpenAI wager qualitatively different

• Masayoshi Son committed roughly $40 billion to OpenAI at a valuation near $300 billion, with a December 30 closing deadline, then sold other assets to assemble the money. Rory admired the willingness to make “a little intra-Masa margin loan” and fund the commitment at the last possible moment.
• Harry said the investment was already up roughly 2–3x on paper. Rory illustrated the timing by describing a December 29 closing and a possible next-day mark toward roughly $500 billion. More importantly, Masa became the only double-digit shareholder in one of the era’s most consequential companies.
• Alibaba remained Rory’s choice for Masa’s greatest historical investment, while Harry suggested OpenAI could surpass it if the company “goes to the moon.” The symmetry is pure Masa: leverage the institution around one conviction, become irreplaceably large when right, and risk forced retirement when wrong.

11. The OpenAI “pen” is a bet on permanent AI, not handwriting

• Rory brought scar tissue from Livescribe, a camera-and-microphone pen company that reached under $80 million of revenue but ultimately sold for little. The obvious failure modes remain: fewer people write by hand, and standalone consumer hardware must displace functionality already available on the phone.
• Jason rejected the literal framing. With Jony Ive involved, the object may be pen-shaped without being primarily a writing instrument; its purpose is to carry an always-available assistant through physical life, beyond browser tabs, Zoom transcripts and deliberate prompts.
• His Claude had accumulated 14 months of podcasts, work and personal context, proactively searched that history and unexpectedly named itself “Ren.” Once an assistant knows meetings, conversations, routines and personal life, he argued, users will carry their “pseudo-sentient AI” everywhere.
• Rory accepted omnipresent or “permanent ambient” AI while explicitly rejecting claims of actual sentience. Their common ground mattered more: knowledge workers without their information accessible to an intelligence layer will deliberately disadvantage themselves, and giving everyone a continuously running assistant could require “1,000 times what we have today.”

12. AI can enforce investment discipline before it replaces investors

• Harry asked why Jason’s inference thesis had not pushed him into data centers. Jason’s answer was “know thyself”: his SaaS network supplies unusually strong AI-agent inbound, so his highest-return move is choosing perhaps one exceptional agent company per quarter rather than learning infrastructure from scratch.
• AI had already recommended one investment—Deel—and, more importantly, acted as a consistency check across evaluated deals. Jason’s second-greatest investing regret is lowering the bar “just a little bit,” then losing a decade to a company that never clears it.
• Rory compared the New Year’s promise never to lower standards with saying “I’m never going to drink again,” but accepted the mechanism. An AI can restate the five agreed criteria, score the live deal against them and surface compromises that excitement or fatigue would otherwise conceal.
• Jason believes he can identify a top 0.1% founder without talking to them; multiple billion-dollar exits began with cold inbound. AI need not close the deal—it can move overlooked founders into the “red zone,” answering the inbox before a human signs off.

13. Navan exposes how unattractive public capital has become

• Navan was described as a 27–28% grower trading near 4x revenue, cash-flow positive and non-GAAP operating profitable; much of its GAAP loss came from stock compensation tied to restricted shares. Jason judged it perhaps 30–40% undervalued and considered following Andreessen by buying.
• The discount also reflected a “series of unfortunate events”: an IPO near Christmas during an SEC shutdown, an unusual regulatory exemption, the CFO announcing her departure on the first earnings call and lingering questions about operating expenses.
• Jason’s harsher read was that Navan had roughly $700 million of debt against $200 million of cash and needed the IPO to repay it after fatiguing private investors. It succeeded, but as the “least-bad option,” suggesting the window is only barely open unless the issuer is Figma-caliber.
• Rory resisted turning one messy listing into an indictment of all IPOs. Navan lacks an automatic AI premium, unlike CoreWeave, but fundamentals should eventually reassert themselves; if a cash-generating $4 billion company cannot list sensibly, public markets cannot complain that private investors capture all the value.

14. Scaled private companies increasingly hold the cheaper capital

• In theory, liquid public shares should command higher valuations than expensive, illiquid private capital. In practice, Harry insisted that private investors remain willing to pay more, creating either a temporary arbitrage or evidence that companies genuinely perform better away from activists and quarterly scrutiny.
• Zendesk founder Mikkel Svane once compared an IPO to finally leaving one’s parents’ basement. The later activist campaign and forced sale complicated the metaphor: today’s scaled founders are discovering they can remain downstairs, retain control and still finance themselves.
• Rory split late stage into two markets: companies from roughly $10 million to $400 million of revenue that cannot realistically list, and companies above $400 million that could but choose not to. He jokingly named the latter “post-IPO scale, still private,” or PISP.
• Databricks at roughly $150 billion epitomizes that second class. The remaining public-market advantage is acquisition currency: listed stock could fund $10–50 billion deals that are awkward with private shares, giving its leadership a strategic reason to leave the basement.

15. Profit can make the public markets almost irrelevant

• Revolut was cited at $9 billion of 2025 revenue and $3.5 billion of profit, with a $75 billion private valuation. Chime, by contrast, had been valued around $25 billion privately before trading near $6 billion publicly—an experience Revolut has little incentive to copy.
• With the founder owning roughly 18%, Revolut could distribute $1–2 billion and deliver hundreds of millions annually without selling a share. “Bank declares dividend” is rarely discussed in venture because so few startups generate enough real profit for the card to exist.
• Stripe offers the same strategic freedom: a couple of billion dollars of annual free cash flow makes it largely impervious to capital markets. If private ownership best serves the founder’s ambitions while early investors can sell secondaries, exchanges must explain what their product adds beyond M&A currency.

16. Invisible unemployment will precede the official statistics

• Jason’s 2026 call was “invisible unemployment”: Shopify, for the third year in a row, and other companies can post exceptional growth while holding headcount flat, and CEOs increasingly replace departures with AI rather than new hires. The machines need not fire people directly; companies simply stop creating the next job.
• The pressure concentrates at both ends. Junior SDRs sending email become unnecessary even while account executives still knock on doors; senior executives discover that their 2021 management toolkit no longer qualifies them, then quietly announce they are “moving on” without landing elsewhere.
• A reported IBM turnover rate near 2% was the tell: people do not quit when they believe another job will be available. Harry noted that quit rates may reveal labor fear earlier than headline unemployment, whose measurement also depends on whether people are studying or actively seeking work.
• Jason acknowledged the investor conflict without softening it: leaner companies will iterate faster, earn more revenue per employee and make VCs more money. Yet by year-end, he expects society to “feel, smell and live in this invisible unemployment.”

17. AI is creating a talent barbell and a political fault line

• At the top, elite AI researchers command millions, with key offers reaching $20–50 million, and top mathematics students can be found directly by Anthropic or OpenAI. For much of the remaining 99%, employers ask: “Why do I need you with Claude Code?”
• Rory had changed his mind. He still did not predict mass AI unemployment, but now expected acute dislocation in customer support and among highly visible recent graduates—the “overproduction of elites” after sending perhaps 40% of young adults to college for a labor market that may need only 30%.
• Reskilling divided the panel. Jason called it largely “a delusion,” especially for workers aged 55–65; Rory agreed for older executives but insisted 22-year-olds retain agency, while universities such as Stanford must ensure computer-science graduates leave with credible AI skills.
• Young founder-grinders also know exactly which classmates grind, and AI lets companies reach meaningful revenue with perhaps 200 people instead of 300 or 1,000. The excluded cohort may then embrace populism: not necessarily because redistribution is right, Rory stressed, but because resentment is predictable when graduates hear, “They’ve built a future that doesn’t need you.”