NVDA GTC, M&A Wiz / Goog $32 B Deal, April 2 Tariff Uncertainty; Huawei Belt & Road; ChatGPT | BG2

Summary

• Brad’s February “golden age of uncertainty” call has played out — NASDAQ off ~10%, many components down 20-30% — and the market may now be rolling into leg two: recession fear, which “can become a self-fulfilling prophecy.” The data is genuinely mixed: consumer and business confidence worsening, airlines (Delta, United, Frontier) flagging demand, GDPNow turning down, and the Fed just cut its GDP median to 1.7 from 2.1 while raising inflation to 2.7 — yet BofA’s real-time card data shows consumer spend up 6% y/y and hedge funds sit in bottom-quartile exposure wanting back in.
• On April 2 tariffs, Brad is emphatic this is not a negotiating tactic — the administration has “a principled view about restructuring globalism” (JD Vance’s “two conceits of globalism,” Bessent’s “American detox period”). His tradeable frame: tariff revenue going from $65B to ~$150B is easily absorbed; $65B to $1 trillion “pushes us into a recession.” Expect no big flows into the market for 60 days while participants wait for the clearing event.
• Google’s $32B all-cash Wiz deal (~30x+ forward revenue on ~$1B ARR, vs Google at 5x) is the litmus test for whether M&A animal spirits get unleashed. Bill’s read: EU pressure blocks Google from buying anything in search, AI, or YouTube, so enterprise security was a plausible lane left — and Ferguson’s FTC (“I’m not a regulator… I’m a cop on the beat,” “go to court or get out of the way”) is a genuine break from Khan-era purgatory. With public software at 6-7x revenue and this deal at 30x, “there’s ample room for buyers and sellers to meet in the middle.”
• At GTC Jensen doubled down: DeepSeek was “profoundly misunderstood,” compute needs are “a hundred times greater than what we believed a year ago,” and the AI data-center TAM reaches $1 trillion per year by 2028 — versus Wall Street consensus that Nvidia somehow tops out at $250B revenue. At 20x next-year earnings (below Costco’s 50x), Brad sees no re-rating needed — just earnings growth in a market where Nvidia holds or gains share.
• Bill’s counterweight on Nvidia: pre-training scaling has “at least slowed materially,” the token-count framing is “an ultra promotional way to talk about things,” and the “chief revenue destroyer” riff creates a real economic problem for customers who just stretched server depreciation from four to six years. His net: of the MAG7, Nvidia is both “executing the fastest” and “most exposed to global trends” — and “the number one risk on the stock is government action from DC.”
• Both agree the Biden AI diffusion rule plus high structural tariffs would be “unilaterally disarming America in the race to AI” — export controls already backfired, with DeepSeek running inference on Huawei’s Ascend 910 and China building a vertically integrated chip supply chain. Bill’s warning: repeat the Nortel telecom mistake and you get a “Huawei Belt and Road” across Saudi, UAE, India, and Southeast Asia.
• Consumer AI may already be decided: ChatGPT holds #1 in the App Store while DeepSeek fell out of the top 100 and Grok slid to ~65 — “you’re going to have to be 10x better than ChatGPT to slow down that inertia.” OpenAI is “massively supply constrained” at 400M+ weekly users, building Abilene and Denton campuses just to launch “four or five products literally sitting on the shelf.”
• Bill’s “Gurley negative gross margin AI theory”: stacked negative-margin layers (hyperscaler → model → app) subsidize end prices and triple-count revenue, VC money plus winner-take-all pricing psychology makes it systemic, and “you could have substantial resets as people get in touch with unit economics.” Brad’s corollary — Anthropic’s AWS-mediated revenue carries “very little or no contribution margin,” while mostly-direct OpenAI is the best-positioned independent; the AI internet analogy should be e-commerce, not Google.

Deep dive

1. The uncertainty gas is out; now comes the recession-fear leg

• Brad’s scorecard on his early-February call: “maximum political uncertainty, maximum economic uncertainty, maximum technological uncertainty” pushed discount rates up and multiples down — NASDAQ -10%, components -20-30% since. “Just in the last week or so we started rolling over to the second potential leg of concern, which is fear of recession and growth” — which “can become a self-fulfilling prophecy.”
• The negative tape: consumer confidence worsening (record-high share seeing business conditions deteriorating), airlines as the bleeding edge (Delta, United, Frontier all flagging demand; TSA passenger growth a leading indicator), GDPNow turning down, and the Fed’s fresh dot plot moving all three dials the wrong way — GDP 2.1→1.7, unemployment 4.3→4.4, inflation 2.5→2.7.
• The offsetting real-time data: Bessent sees no consumer slowdown in card and bank data, and BofA reported spend up 6% y/y on $1.5T of quarterly consumer spending — big-ticket travel is just shifting to local entertainment and restaurants. Meanwhile every manager Brad talks to is in “their bottom quartile of exposures” but believes the back half and next year “is going to be really great.”
• CEOs, though, “are absolutely more on hold today than they were before” — they want to see what April 2 brings before making big decisions.

2. April 2 is a beginning, not an end — and the dollar size is everything

• Brad’s strong view: “This is not a negotiating tactic by Trump… they have a fundamental view that the middle class was hollowed out by effectively sending all of this labor offshore.” Vance’s “two conceits of globalism”: the belief that America keeps the high-value work while China does low-value (instead BYD designs better cars, likely CATL a better battery, and DeepSeek runs inference on Huawei’s Ascend 910), and that “cheap labor is fundamentally a crutch to innovation.”
• Mechanics: roughly 15 countries — Brad thinks he heard Bessent call them the “terrible 15” — get country-by-country tallies of tariff and non-tariff barriers on April 2; some are pre-negotiating. “I don’t see how that can be the end. It feels like a lot closer to the beginning of a negotiation.”
• The quantification that matters: last year’s US tariff revenue was ~$65B. Going to ~$150B “by just making it more fair across the board” is easily absorbed; targeting $1 trillion “pushes us into a recession.” Bessent’s stated logic — remove your barriers and ours come down — “if that’s the objective… it’s a good policy and one the economy will easily absorb.”

3. Bill’s dissent: comparative advantage, and don’t provoke Europe into China’s arms

• Bill’s principled pushback: “I remain, based on all the learning I’ve done to this point in my life, a big believer in comparative advantage… water runs downhill and that’s where those jobs want to go. I don’t think there’s anybody that wants to build a $40 microwave in America.” China has, Bill suggests, roughly 35% of the global auto market while producing better cars faster — “does anyone really expect GM and Ford to reorient themselves overnight?” Even full automation “doesn’t bring the employment back.”
• His moral counter to the hollowing-out narrative: it “happened simultaneously with bringing 500 million people out of poverty in China” — and “protecting the US by trying to somehow injure or disable a competitor that’s in a different country in the long run will make us weaker.”
• On style: “if reciprocity is the goal… you can use math and show slides and you don’t have to be combative.” He relays the Economist’s editor’s warning that European leaders are so “angry and chafed” that “they may come to trust China more than the US” — “talk about picking a fight you didn’t want.”
• Brad’s partial concession: the bottom 40% “has not been keeping up,” most CEOs and even a majority of Hill Republicans agree with Bill against high structural tariffs — but there’s a middle ground of fair trade with incentives for structurally critical industries, evidenced by TSMC’s $100B including “an advanced R&D fab in the United States — that’s critical.”

4. Wiz at $32B: Google’s plausible lane, and Brad’s Lacework scar

• Bill’s reverse-engineering of the deal: Google, an acquisition-proven top-five market cap, faces EU pressure that likely blocks buying anything in search, AI, or YouTube — so “which of our businesses would we be allowed to acquire in?” Enterprise, where they lack dominant share, was a plausible lane. Differentiation vs developer-friendly AWS and Microsoft’s corporate leverage: “security, especially in a distributed world, seems like a very reasonable answer.” Even so: a $3.2B breakup fee and closing maybe not until next calendar year — “let’s not count the chickens just yet.”
• The numbers: $32B all-cash, ~$1B ARR rumored this year, 30x+ forward revenue against Google’s own 5x, buying roughly 2% growth on ~$45B of cloud revenue. A company started in 2020, sold five years later — hat tip to likely Shardul Shah at Index, Doug Leone, Greenoaks.
• Brad’s vulnerability, on record: Altimeter backed Lacework with Sutter Hill in a very similar space — “Lacework raised a lot of money, too much money I would argue… they didn’t have the product-market fit and execution relative to Wiz. I remember when Wiz did that Series B and I said, we’ve got a product issue — they’re killing us.” The lesson: “it is a razor’s edge between the promised land and not making anything.”
• Bill’s takeaway his partners dispute but he’ll say anyway: “repeat entrepreneurs in the enterprise space are a golden ticket” — Workday as picture-proof (the team that likely built PeopleSoft on a new platform), Brett Taylor with Fenton, likely Slootman — because each stack transition (mainframe→mini→client-server→internet→cloud) lets the experienced operator run the playbook again.

5. Ferguson’s FTC is not Khan 2.0 — and the pipeline knows it

• Brad read the transcripts behind the scary “cop on the beat for big tech” headlines: Ferguson said “I’ve written hundreds of pages of dissents with Lina Khan,” and his core differentiation is speed — Khan’s FTC “lingered,” creating “a regulatory fear state” where actions died on the vine. Ferguson: “I’m going to go to court or get out of the way and let business thrive,” and “I’m not a regulator… I’m looking for fraud, monopoly, or collusion.” Bill: Google/Wiz “passes that test — it’s none of those three things.”
• The arbitrage: great public businesses growing 10-30% trade at 6-7x revenue while Wiz cleared 30x — “ample room for buyers and sellers to meet in the middle if the United States is truly open for business.”
• The wrinkle: Europe. If the EU keeps “biting our ankles” on M&A between two American companies, Brad expects that to get “kicked into the domain of trade negotiations” on April 2 as a non-tariff barrier.
• IPO front, from the buy side: “our phones are ringing” — CoreWeave in the chute, rumors about likely Klarna, and Brad thinks likely Cerebras is in the chute, plus “many others.” With the NASDAQ down 10% in a month and M&A activity alongside a full IPO pipeline, “this to me is a strong signal that people believe this is much more of a pro-growth administration.”

6. GTC: Jensen re-doubles the TAM while trading at 20x

• Setup: stock peaked ~$150, touched $105, back to ~$115 — ~20x next-year and ~24x this-year consensus earnings, below the S&P, while “Costco’s trading at 50 times.” The worry wall: is AI overhyped, what about DeepSeek, what about ASICs.
• Jensen on stage: “a profound and deep misunderstanding of DeepSeek R1” — compute needed is “a hundred times greater than what we believed a year ago.” The TAM slide has been re-raised twice: from ~$250B/yr a year ago, to ~$500B/yr, now to $1 trillion per year by 2028 in new AI workloads plus x86 data centers rebuilt as accelerated compute. Driver: “by the end of the year most code is going to either be refactored or written from the start using machine learning” — coding agents run on accelerated compute, echoed by Zuckerberg, Musk, and Altman.
• Breadth and cadence: robotics and autonomous already >$5B run rate, Blackwell 40x Hopper on TCO, top-4 CSP orders of 3.6M Blackwell GPUs this year vs 1.3M Hoppers in the comparable period, a photonics leap, and — Brad’s Apple comparison — a public four-generation roadmap “Apple would never do.” On ASICs, the line that “will roil some people”: “the vast majority of custom ASICs that get designed and taped out actually never scale into any commercial production” — and Trainium is cutting prices, which “you usually don’t do when things are going great. I think Trainium’s not even competitive with Hopper.”
• Brad’s framework: multiple compression is done — “you’ve seen all the multiple compression you’re likely to see” — so returns track earnings growth; consensus implying Nvidia tops out at $250B revenue requires a share collapse he doesn’t believe, “if anything they probably gain share.”

7. Bill’s GTC skepticism: scaling nuance, promotional tokens, and the depreciation problem

• On scaling: “most people believe pre-training scaling, if it hasn’t topped out, has at least slowed materially” — even Nvidia talks about inference being 99:1 — and Jensen’s “everyone got it wrong” framing is “the bulls not wanting to admit that anything isn’t scaling.” On token counts: nobody in Intel’s heyday counted CPU clocks — “I do find it kind of an ultra promotional way to talk about things.”
• On likely Rubin (disclosed bias: Benchmark is an investor in likely Cerebras): more on a bigger die “is the direction of Cerebras… that one confirms to me we’re moving more towards inference.”
• The “chief revenue destroyer” worry: customers just moved from four-year to six-year server depreciation — “if he’s going to kill them in two years, there’s an economic problem for these customers.” Brad’s rebuttal: CSPs elongate depreciation because software upgrades keep flowing to Hopper — “it’s not that the old chip wasn’t the right thing to buy at that moment in time,” Hopper remains great for inference.
• Bill’s concession, unprompted: ranking the MAG7, “which is executing the fastest right now, and which is most exposed to global trends? Nvidia would win both.” Brad, position transparency: Nvidia is smaller than prior years because all his sizes are down, and he’s assuming flat multiples — pure earnings-growth compounding.

8. Diffusion rule + tariffs = “unilaterally disarming America” — the Huawei Belt and Road

• Bill’s flat-out call: “the number one risk on the stock is government action from DC” — Singapore billing-vs-shipping questions, DeepSeek training location angst — and it’s bipartisan and “a lot of it’s misplaced and overly angry.” His jab at OpenAI’s anti-DeepSeek paper: “quite sad… if GM or Ford put out a paper like that about BYD, we would look at them and go, oh, you just want the government’s help. You must be uncompetitive.”
• Bill’s thesis: high structural tariffs plus the Biden diffusion rule is “unilaterally disarming America in the race to AI.” Export controls already backfired — DeepSeek runs inference on Huawei 910s (“they just throw a hell of a lot more power at these chips”), and restrictions forced China to build a vertically integrated design-to-fab supply chain. Frontier models are out of the barn: DeepSeek, Alibaba, ByteDance all have them.
• The historical rhyme: “they said this would never occur in telecom equipment because of Nortel… Huawei ran the table across the globe because the US did this. We ought not make the same mistake twice.” The diffusion rule blocks chips to Saudi, UAE, India, Southeast Asia — “I just see Huawei walking right through that door.” Bill adds ASML: a Dutch company at 80-90% share, visibly chafed at US extraterritorial authority — “I could easily see the ASML team just say, screw it, we’re selling to China.”
• Brad’s middle path: conditional tariffs — announce them but suspend for two years pending proof TSMC and Nvidia are onshoring, “we want to provide the incentive, but we don’t want to slow down US companies in the race to AI.” He disagrees with Bill’s Morris Chang skepticism: with today’s fab automation and the strategic stakes, the Arizona pressure “is achieving the desired results.”

9. Consumer AI: ChatGPT’s inertia may already be insurmountable

• The App Store tape: DeepSeek spiked to #2, now out of the top 100; Grok hit #2, now ~65; Gemini “never really jumped,” ~55; ChatGPT still #1. Brad’s pattern recognition from Google and Meta: “these tend to be winner-take-most markets… you’re going to have to be 10x better than ChatGPT to slow down that inertia.”
• OpenAI at 400M+ weekly users is “massively supply constrained” — building the Abilene (Oracle) and Denton (CoreWeave) campuses “just to actually launch the products they currently have,” with “four or five products literally sitting on the shelf.” Expect “pretty profound” releases in the next 8-10 weeks.
• Bill can’t take the other side — “I don’t have any ability to take the other side of your argument” — but flags the counterweight: “Netscape had Microsoft; OpenAI has four or five of the most powerful companies in the world aimed at them,” and crushing voice might be the wedge. Both are struck by the dogs that haven’t barked: Meta “in full beast mode but unusually quiet” on a standalone consumer app, and Apple showing “no real change.”
• Bill’s open-source provocation: DeepSeek is more open than Llama — no usage caps, forked ~1,500 times on Hugging Face, prevalent in the enterprise — so this is “an interesting moment for Meta to consider going left and getting even more open,” and “even OpenAI could consider it”: if consumer inertia is the moat, open-sourcing “would undermine anyone else trying to compete on the quality of their proprietary model.”

10. The Gurley negative-gross-margin AI theory

• The setup Brad frames: Anthropic reportedly has “very little or no contribution margin” because most enterprise API revenue routes through AWS with a “huge customer acquisition fee” — TAC, in Google terms — plus serving costs; mostly-direct OpenAI avoids that. And contribution margin ignores overhead and training: “if you’re not meaningfully positive at contribution margin, you have no chance of covering the cost on a fully burdened P&L.”
• Bill’s “first draft” theory: stacking. A voice startup on a model provider on a hyperscaler — “if those second and third layers are both negative gross margin, the consumer is buying compute at a price lower than if they bought it directly… and you’re literally triple counting revenues.” Add VC billions, winner-take-all psychology (“you can’t lose market share, so how do you price?”), platform credits muddying capex, and financially unsophisticated founders: “it’s set up to be messy, and you could have substantial resets as people get in touch with unit economics” — on “a very unpredictable time window.”
• His three closing riders: steep model-price declines make negative-margin pricing feel rational (“in 6 months the compute for this will be a lot cheaper”); production users optimize to cheap — which is why open source and DeepSeek thrive in the enterprise; and the right internet analogy is e-commerce “selling dollars for 85 cents,” not Google. Hedged as promised: “I’m not suggesting it’s 100% true. It’s possible this is going on.”
• Brad’s contrast: Google raised under $50M before IPO with near-zero marginal cost to serve — the AI winners will be those who “fully control the top of the funnel” with no marketing cost. But there’s no clearing event coming: Meta has “a printing press kicking out billion-dollar bills in the back room,” Google and Amazon likewise, and Elon/Grok has “a unique ability to raise global capital — you absolutely can’t count him out.” The next 90 days — product releases, April 2, reconciliation — decide whether it’s “off to the races come this summer” or “the middle of a trade war.”