The Impact of H1B Visas on Startups in the US & NVIDIA Invests $100BN Into OpenAI
Summary
- The Nvidia–OpenAI financing loop buys permission to test the scaling thesis; it does not prove the economics. Capital availability means nobody calls timeout for another year or two, even if the marginal $300 billion ultimately earns little: “We will find out.” Another panelist reported Sam Altman saying OpenAI needs three orders of magnitude more compute, making heroic revenue projections the load-bearing assumption.
- Nvidia’s $4.5 trillion valuation rests on startling customer concentration, but those customers are still “determined to spend themselves into oblivion.” Roughly six buyers reportedly generate 83% of revenue, versus billions of users at Apple or hundreds of thousands of enterprise customers at Microsoft. That single-threaded exposure is dangerous, yet OpenAI, Google, Meta and Oracle show no sign of blinking.
- The tradeable phenomenon is an AI capex boom running far ahead of the AI revenue base. The panel contrasted roughly $600 billion of annual capex with only $30 billion–$40 billion of current revenue; data labelers and infrastructure vendors thrive because the six major buyers optimize for speed, not price. The closest rhyme is 1999, except balance sheets, vendor equity and capacity guarantees can sustain this cycle longer.
- Venture has split into a familiar early-stage market and a separate ultra-late-stage private-public market. Seventy-five percent of 2025 VC dollars went to 19 companies, but the remaining 25% still funds roughly the same Series A ecosystem. Below the obvious “S tier,” funding has become murky: triple-triple-double-double still works at meaningful scale, though meetings and conviction take far more effort.
- Returns depend more on entry price, ownership and bet sizing than on owning the era’s most prestigious company. Early OpenAI investors and a Series A investor in Netskope might each show roughly a 7x blended return—“all 7Xs are exactly the same because money is fungible”—despite radically different company significance. The real decision is whether to crystallize a fund returner or let it compound toward 2x while accepting concentrated downside.
- Navan’s IPO case shows why a credible number-three player may rationally go public before cleaner comparables. Its filing showed $613 million of revenue, 32% growth, 10,000 customers and 110% NDR; flat-to-down operating expense suggests a hard push toward profitability. Going first can secure novelty, liquidity and acquisition currency before Ramp or Brex makes public investors ask, “Why do I need your IPO?”
- The market is repeating 2021’s behavioral errors even as mature software companies finally confront 2021’s marks. Notion’s reported $500 million ARR and reacceleration show AI can revive scaled SaaS, but the panel’s rough public-market discussion put a 30%–40% grower nearer a $4 billion–$5 billion range than old private expectations, depending on revenue and growth. Meanwhile hot AI rounds can close on a Saturday with effectively no diligence, making “founder honest” more meaningful than the now-hollow claim of being founder-friendly.
Deep dive
1. OpenAI has secured the capital to run the scaling experiment
A panelist rejected the “infinite money-printing machine” framing around Nvidia investing in OpenAI, OpenAI committing $300 billion to Oracle and Oracle buying Nvidia chips. Like any aggressive financing structure, everyone looks brilliant if the underlying business works; if it does not, “it all comes back and bites you in the ass.”
The consequential fact is that Sam Altman gets to apply enormous capital without anyone calling timeout for perhaps another year or two. If OpenAI’s projections of more than $100 billion in revenue are real, capital and chip access will let the market discover that; if the marginal $300 billion earns no return, the same experiment exposes it.
Harry challenged the premise: GPT-5’s emphasis on efficiency and less dramatic gains appeared to show scaling laws already weakening. Another panelist’s rebuttal was behavioral evidence: Altman said beside Jensen Huang and Brockman that this is “just a start” and OpenAI needs three orders of magnitude more compute, not merely 10 times more.
The first panelist did not endorse that forecast. The level of “heroic assumptions” required is high, but OpenAI has been astonishingly right for six years, so backers will keep doubling down “until the return on the double-down isn’t there.”
2. OpenAI’s advantage is momentum, while Nvidia owns the sharper monopoly risk
Asked whether Anthropic should feel structurally disadvantaged, a panelist separated optics from constraints. Anthropic was reportedly turning investors away and could likely raise another $10 billion immediately, so capital itself is not scarce; preferential GPU access might matter, but the concrete problem solved by OpenAI’s deal remains unclear.
Another panelist argued that OpenAI’s deeper advantage is optionality: enough capital to develop its own GPU and secure hundreds of gigawatts of compute. The discussion also suggested that Altman carefully avoids appearing monopolistic, preserving credible rivals and rarely denigrating them because ChatGPT already approaches Google Chrome-like consumer share.
The monopoly critique was met with a cynical counterpoint: classical monopoly harm comes from extracting excess profit, yet ChatGPT presently subsidizes consumer surplus estimated in the tens of billions. It has commanding chatbot share, but Gemini and Perplexity weaken the literal monopoly claim; Nvidia is closer to monopoly market share and faces customers actively designing substitutes.
The unanswered strategic question is whether Nvidia restricted OpenAI’s chip ambitions when taking equity. Giving capital to one of six dominant customers while that customer builds a competing processor would be unusual, but any such agreement is precisely the detail neither side is likely to highlight.
3. Six buyers make Nvidia both extraordinarily powerful and unusually fragile
The panel cited roughly six customers accounting for 83% of Nvidia’s quarterly revenue. A company valued around $4 trillion–$4.5 trillion therefore depends on the spending decisions of six or seven people, unlike Apple’s roughly two billion users or Microsoft’s vast enterprise base.
The framing captured both tails: “You’ve got this company, what, four and a half trillion, with only six customers. That’s bad news.” The offset is that OpenAI, Google, Meta and Oracle have all signaled that they will not blink.
Nvidia consequently has elements of both monopoly and monopsony exposure: it dominates a critical input but sells disproportionately to a tiny group capable of developing alternatives. Its near-term defense is that every member of that group is “determined to spend themselves into oblivion to win the prize.”
Concentration propagates downstream. Mercor reportedly gets 55% of revenue from two customers, and the other four main data-labeling providers described essentially the same pair; those buyers are “incredibly promiscuous” across vendors because they need more supply than any one provider can furnish.
4. The AI capex boom is much larger than the current AI revenue pool
The panel distinguished the AI revolution in applications from the AI capex boom. Adoption is real, but the extraordinary feature is willingness to spend, in aggregate, roughly $600 billion annually against a market presently producing only about $30 billion–$40 billion of revenue, depending on how revenue is counted.
Vendors attached to that spend have prospered because the six buyers are not optimizing cost. A services business with six customers should theoretically be squeezed on price; in practice, data-labeling providers can “stuff” dollars into the bucket because customers care primarily about building faster.
The closest historical rhyme is 1999: unlimited possibility, vendor financing and infrastructure spending well ahead of demand. Nortel and Lucent financed bandwidth customers just as Nvidia now supports buyers, although Nvidia is using equity; the internet thesis proved right, but markets still suffered a five-to-seven-year retrenchment.
The contrast with Web 1.0 was financing durability. Amazon nearly exhausted its cash after its IPO, while today’s leaders guarantee one another’s demand and capacity; CoreWeave might once have failed within months, whereas Nvidia has agreed to buy 300 years of its capacity.
5. Nvidia’s cash machine does not make indiscriminate buybacks harmless
Nvidia’s free cash flow moved from approximately $3.8 billion in fiscal 2023 to $27 billion in fiscal 2024 and $60 billion in fiscal 2025, with roughly $100 billion or so suggested for the following fiscal year. That cash generation gives it unusual capacity to reinvest across the ecosystem.
A panelist nevertheless questioned buying back $9 billion of stock in one quarter near $180 per share and authorizing a $60 billion program. Nvidia’s approximately $60 billion of cash sounds huge but represents only around 1.5%–2% of market capitalization if the cycle reverses.
Another panelist explained the conventional logic: companies often repurchase approximately enough shares to offset employee RSU dilution, keeping EPS and share count constant. The policy was called “as dumb as rocks”—buy when the stock is cheap and retain cash when it is dear, irrespective of dilution.
The broader warning is cyclical. Bull markets direct attention to income statements; difficult markets suddenly make balance sheets valuable, and corporations historically repurchase at peaks rather than lows. Larry Ellison was cited as the counterexample who bought aggressively when Oracle was cheap, then redeployed capital to reshape the game.
6. Expensive markets lower long-term returns before they predict a crash
Harry reported 27 green positions among 28 holdings and an S&P 500 apparently heading toward 7,000, while admitting, “I’m humble enough to know I’m not that good.” The response was more alarming: the speaker was at “0% cash,” recalling 2008 when he had to sell stock down 60%–70% merely to repair his roof.
The panel separated horizons. Starting valuation has weak correlation with one-year returns, so expensive stocks can keep rising as the Fed cuts; its correlation with 10-year returns is much stronger, implying materially below-average prospective returns from today’s entry price.
Holding cash is therefore the accepted price of sleeping well, not a tactical claim that the peak has arrived. Asset allocation should reflect medium-term needs and risk tolerance rather than maximize performance in the current bull phase.
The social froth indicators were telling: LPs, normally hidden “behind the Wizard of Oz curtain,” had begun boasting about returns on LinkedIn. Harry added that public declarations that triple-triple-double-double is dead feel similarly top-like.
7. Venture’s headline concentration masks two different businesses
Seventy-five percent of 2025 venture dollars reportedly went to 19 companies. The panel’s reframing was that the remaining 25% is approximately the same Series A business that has existed for 10–15 years, including roughly 1,000-plus Series As annually; an ultra-late-stage private-public market has simply been layered on top.
Concentration naturally increases by round because companies drop out at every stage. If private lifetimes keep extending, the limiting case is a handful of foundation models plus companies such as Databricks and Stripe raising increasingly late alphabet rounds.
For early investors, the practical underwriting rule remains the next 18–24 months: can the company execute and raise at a risk-adjusted 2x–3x step-up? Forecasting the ultimate five-year value is less useful than identifying the next financing milestone.
The layer immediately below the obvious S-tier is “very, very murky.” Concerns over churn, margins or forward-deployed labor can repel one investor while another preempts at an exceptional price.
8. Strong growth still funds companies, but scale and belief now matter more
The panel rejected the absolutist claim that triple-triple-double-double no longer works. At $10 million–$20 million of revenue, even 100% growth now requires substantially more meetings than 24 months ago; at $50 million–$100 million with triple-digit growth, investors will take the meeting regardless of whether the category is fashionable.
Unloved verticals still face a higher burden. A small, undifferentiated restaurant SaaS vendor may struggle, but an outlier such as Owner demonstrates that exceptional numbers can overpower category prejudice.
When growth is truly outlying, investors may not examine whether revenue depends on forward-deployed engineers or agents. The panel conceded that belief in future durability matters, but argued that Twitter’s hard boundary is overstated—recent IPOs growing around 30% prove there is more than one viable trajectory.
Fund size changes what counts. ICONIQ could celebrate Netskope’s roughly $8.5 billion outcome and Atlassian’s $1 billion purchase of DX, but very large funds increasingly need a few positions capable of returning enormous absolute sums.
9. A prestigious company and a conventional exit can produce the same multiple
The discussion compared early OpenAI investors’ roughly seven-to-eight-times return with what a Series A investor might earn in Netskope. OpenAI is vastly more consequential and reportedly reaches 10% of the world’s adult population weekly, yet “all 7Xs are exactly the same because money is fungible.”
The blended math matters. An investor might earn 25x–30x on first money, 3x on a late follow-on and roughly 7x across $150 million invested; a $5 billion-plus IPO can likewise generate an excellent 10x when ownership and entry price are right.
Absolute capacity is the differentiator. There are few ways to earn a strong return on a very large dollar commitment, forcing the biggest funds toward five or seven giant positions; smaller outcomes still produce excellent equity returns for appropriately sized vehicles.
Asked whether he would sell OpenAI around a $500 billion valuation, a panelist said failing even to consider it would mean “you’re probably just not thinking.” Yet waiting is seductive: the position can double without another meeting, while selling creates taxes and the career risk of missing the next card.
10. The correct sell decision combines valuation with marginal utility
The panel prescribed two steps: estimate fair value and upside from fundamentals, then overlay personal and institutional constraints. “It’s always doubled” is not an analysis when half a trillion dollars already makes a company one of the world’s largest.
Drawing on The Missing Billionaires, the discussion argued that families and institutions usually fail through bet sizing, not stock selection. If someone without $5 million suddenly owns a liquid $5 million position, taking it may be rational even when expected value says hold; risk aversion changes with net worth.
Success itself creates investment advantage. Established firms believe another opportunity will arrive tomorrow, while emerging managers may need immediate DPI; an early win both improves referrals and gives investors “the stomach to roll the dice,” helping explain why early success correlates with later success.
Concentration still requires a threshold, not a feeling of bravery. The cited exercise suggested an investor would need about 70% confidence in Tesla outperforming the S&P to justify being 100% Tesla from its 2010 IPO; the same logic should govern putting 20% of a venture fund into one company.
11. Navan is using IPO timing as both financing and competitive strategy
Navan’s filing reported $613 million of revenue, 32% year-over-year growth, 10,000 customers and 110% net dollar retention. Its survival is part of the story: travel revenue probably went to zero in March 2020, while Oren reportedly carried substantial exposure across vehicles.
The discussion distinguished Navan’s travel-booking economics from Brex and Ramp’s card-led models and BILL’s accounts-payable base. Yet Navan’s post-TripActions rebrand and S-1 claim broader software and payments territory, ensuring public investors will still compare them.
A credible number-three company may want to list before numbers one and two. Going first offers novelty; going last means investors who already own Ramp and Brex can ask why they should attend another roadshow without a major discount.
Flat-to-slightly-down operating expense despite inflation suggests Navan is straining to become profitable, perhaps still one or two years away. Listing now accepts a haircut but secures liquidity, ongoing capital access and public acquisition currency in a receptive market.
12. An IPO begins an 18-month liquidity process rather than ending it
A typical lockup lasts six months, sometimes ending earlier when performance triggers are met. If shares price at $14, trade to $18–$19 and the company delivers its first quarter, insiders may complete a registered secondary; if the stock falls below issue price, that route becomes difficult.
After lockup, a fund can sell or distribute shares to LPs, but board membership brings reporting obligations and narrow trading windows. The panel’s practical estimate was 18 months from IPO to exit, sometimes 24, with directors often leaving after 12–18 months.
Holding while possessing inside information was described as legal; selling on it was not. A director may know that M&A talks could yield a 30%–40% premium and remain silent while LPs demand a sale, but negative information symmetrically prevents selling.
The extreme counterexample is Nvidia’s 1997 venture directors, Mark Stevens and Tench Coxe of Sutter Hill, who have remained on the public board to this day. Stevens was described as possibly never selling a share. The compounding was extraordinary, though portfolio theory would still challenge the concentration.
13. Immigration friction and 2021 marks expose the ecosystem’s constraints
The announced $100,000 fee for new H-1B visas was judged clearly negative at the margin but perhaps modest if it does not expand. The panel cited roughly 440,000 applications, 70,000–75,000 acceptances and estimates of $19 billion–$120 billion in GDP contribution.
One panelist said his first materials-science startup could not have existed without two H-1B transfers among its first 10 employees. Startups may shift founders toward O-1 visas and big technology companies will pay, but the preferred policy is a rational skills-based filter rather than a crude dollar proxy.
Notion’s reported $500 million ARR and acceleration impressed because reacceleration at scale is rare. The valuation discussion put a 30%–40% grower in a rough $4 billion–$5 billion public range, depending on revenue and growth—far below the inherited $10 billion mark but still a real IPO outcome.
The panel’s prescription for Airtable, Notion and other 2021 “decacorns” was to price them on fundamentals—growth, revenue multiples and eventually free cash flow. A proposed deadline was that after January 1, 2026, investors should stop invoking 2021 valuations, even as the panel warned that investors were making the same mistakes again.
14. Hot AI rounds have hollowed out diligence and “founder-friendly”
Harry described investors issuing term sheets to secure exclusivity, conducting real work during a 30-day close and then withdrawing. A panelist called the practice bad but expressed more empathy in 2025: when founders offer one Saturday, minimal data and paid pilots, post-signing discoveries can legitimately break trust.
The alternative is to decline timelines that preclude diligence and arrive with a preformed thesis. The counterpoint was that the best founders deliberately distribute breadcrumbs, then compress the formal decision into one day; aggressive investors will eventually break their stated rules.
“Founder-friendly has become bullshit” was the formulation offered by one panelist. It means writing the check when nobody else will, attending the bad board meeting, recruiting the executive and supporting the company during an SVB weekend—not reflexively saying “great job.”
The preferred term was “founder honest”: tell founders what you actually think, because behavior in a bull market reveals little. Harry recalled a VC wiring personal money during the SVB weekend, reinforcing that tough deals identify both founder-friendly investors and the VCs with whom colleagues genuinely want to work.
15. The quickfire exposed sharp disagreement on consumer hardware
Rory guessed a final US–China TikTok deal might never arrive because dangling it creates continuing leverage; James Gibson chose roughly 60 days, expecting broader China and India tariff negotiations to resolve during the calendar year and some H-1B disruption to dissipate with them.
James Gibson assigned Meta’s new smart glasses “0% chance” of success despite owning six to eight earlier pairs: “We just don’t need to play Tron in our eyes.” Harry disagreed, arguing glasses could unify computing, vision and ordinary life; James reserved more optimism for Jony Ive’s device while acknowledging that changing the phone paradigm is extraordinarily hard.
James judged Atlassian’s acquisition spree defensive rather than transformational. DX and smaller purchases may help move existing customers into AI-enabled engineering management and preserve roughly 20% growth, but the deals will not make Atlassian the AI-dominant coding-agent company. The lonely PR image of Michael Cannon-Brookes on the DX deal was offered as a sign of how difficult the transition is.