Liberation Day, Tariffs, US v China Open Source, OpenAI, $CRWV, TikTok | BG2

Summary

• Liberation Day landed at the big end: headline tariffs jump from $77B last year to ~$750B — above even hawk Peter Navarro’s $600B — with China at a combined 54%, a 10% floor on every country, and 25% on autos, though pharma and semiconductors were exempted (Brad’s hunch: effective number “right around 600 billion”). S&P/NASDAQ futures swung 600bps lower during the presentation, on top of indices already down 8-15% YTD.
• Brad’s tradeable call: the plane lands at $300-400B, “not 6 or 700 and certainly not a trillion” — Trump “can’t afford to lose a single Republican vote” on reconciliation and senators are hearing it from constituent CEOs. No clearing event for 30-60 days, and the key one is Xi-Trump: there’s “no way that lands at 54%.” Meanwhile, “the best opportunities to buy something are when people are a little fearful.”
• Gurley’s core objection isn’t the rate, it’s the ambiguity: reshoring a factory might take ~3 years at higher cost — “I don’t think our labor is globally competitive, nor do I think it wants to be” — and not knowing policy 3-12 months out “makes it very hard to allocate capex in any meaningful way whatsoever.” The Montana CEOs, “almost to a one,” said February and March slowed; one had four contracts cancelled by angry foreign counterparties.
• China’s open-source push is longstanding, not a Xi directive — the perfect answer to 40-50 years of IP-theft accusations — and DeepSeek (1,500+ Hugging Face forks) is prolific. DC may move to limit it, opening a short window for a US player to “go left of DeepSeek in terms of openness”; Gurley: “the next 3 to 6 months could dictate who’s standing on top of the hill 5, 10 years from now.”
• OpenAI’s open-weight model with reasoning — one of its first since GPT-2, explicitly without Meta’s 700M-user license cap — reads to Gurley as a Kubernetes-style deft move that “wipes out the business opportunity for other models on the API side.” Llama 4 is rumored this month: 400B-parameter MoE, ~10M context window. Washington’s motive per Brad: no “Huawei-DeepSeek Belt and Road.”
• OpenAI’s $40B SoftBank-led round at ~$260B pre ≈ 20x forward run-rate revenue vs Anthropic ~50x and X/xAI ~80x — the leader priced at a discount to peers. Gurley: headline “smells of being promotional,” it’s a “sport of kings” where unit economics get postponed ($5-7B losses, ~$15-20/yr per free user); Brad’s rebuttal: Uber was “never going to be profitable” in 2020 and does $6B FCF this year — and demand is throttled at 1M new users/hour, 500M WAU, 20M paid.
• Agent economics break Google’s LTV math: Booking.com pays Google $50 on a $20 hotel take using lifetime-value logic; a white-label service under an agent can share at most $10 — all upside for OpenAI, a major replacement risk for Google (already $200→$150). The antitrust irony: “finally in 2025, the first time we’ve actually had competition for Google… now we get around to breaking up their search monopoly.”
• $CRWV broke issue at ~$37, then rallied to ~$60 on a Google/Nvidia Grace Blackwell deal that answered the Microsoft-concentration critique; Brad argues GPUs retain value beyond two years (A100s still working in year five, V100s ran seven), while Gurley questions whether depreciation should be accelerated — “the answer may lie somewhere in between.” And the proposed TikTok fix: a NewCo licensing the algorithm with ByteDance under 20%, unlocking ~$600B of trapped US venture value — unless China, freshly hit with 54% tariffs, blocks it.

Deep dive

1. Liberation Day lands at the hawkish end — ~$750B headline before exemptions

• Brad’s setup from the JPMorgan event in Montana (100 tech CEOs; Howard Lutnick and likely Elon Musk, David Sacks, and Doug Burgum speaking): this policy is “doctrinal” — the only question was whether tariffs landed near the Navarro/Lutnick $600B-$1T end or the Bessent/Hassett low end. Answer: the big end — headline tariffs go from $77B last year to roughly $750B, above even Navarro’s number.
• The tape told it live: a WSJ headline of 10% across the board popped futures ~2.5%, then the reciprocal chart hit — 34% on China on top of the existing 20%, 54% total — and S&P/NASDAQ futures fell 600 basis points from the initial jump, on top of indices already down 8-15% on the year.
• The structure: 25% autos (Mexico, Canada, Germany — UAW members in the front row, “We won the state of Michigan because of this”), reciprocal tariffs effective April 9, a 10% minimum on all countries, China in a bucket of its own. Then the footnote: pharma and semiconductors exempted — Taiwan gets 32% but chips skate — so Brad’s hunch is the effective landing is “right around 600 billion.”
• On the “reciprocal” math itself: non-tariff barriers (currency manipulation, judicial restrictions on US products) fill in the numbers, and Brad is blunt — “you can kind of make those numbers whatever you want to make them.”

2. Uncertainty, not the rate, is what freezes capex

• Gurley’s first principles: a believer in “open markets, free trade, and comparative advantage” — “very solid mathematical arguments” that trade walls hurt both sides. And reshoring isn’t a dial: moving production from Thailand might take ~3 years at higher cost, because “I don’t think our labor is globally competitive, nor do I think it wants to be.”
• The deeper poison is ambiguity — nobody knows what’s negotiation bluster versus policy 3, 6, or 12 months out, “which makes it very hard to allocate capex in any meaningful way whatsoever.” Businesses would probably pause hiring, cascading into unemployment and consumer spending.
• Brad’s ground truth: the CEOs “almost to a one” said things slowed in February and March “because nobody knows what to do”; CEOs were texting him during the speech asking “do you see us exempted?” The Fed had just cut GDP forecasts, raised inflation and unemployment forecasts.
• Cascading effects already visible: one CEO had four contracts cancelled (three European, one Asian) by counterparties upset with America; China, Korea and Japan planning to collaborate in a response — “we haven’t seen them combine forces on anything since the Mongols” — and European leaders heading to China for closer trade ties, the realignment Brad had warned about.

3. Brad’s call: $300-400B, and China is the key clearing event

• With NASDAQ down ~18% peak-to-trough post-Trump and many names off 40-50%, fear is being priced: “the best opportunities to buy something are when people are a little fearful” — Altimeter is getting “a little bit more positive on the companies we like the best.”
• Why moderation is structurally forced: Trump “can’t afford to lose a single Republican vote” on the reconciliation package — the “big, beautiful bill” with no tax on tips and permanent tax cuts — so Brad’s landing zone is $300-400B in tariffs, “not 6 or 700 and certainly not a trillion.”
• The 30-60 day watchlist: do the semi/pharma exemptions hold, country-by-country renegotiations (Vietnam already capitulating pre-deadline), and above all China — “it scares me how big the tariffs are” and Xi-Trump talks are, in his view, imminent. Until then, “we’re still in the fog of war.”

4. China’s open-source embrace is longstanding, not a Xi directive

• Gurley calls the “Xi forced DeepSeek open source” theory “remarkably misplaced”; Brad says Chinese majors have sponsored Linux and major OSS projects for over a decade, because after “40 or 50 years” of IP-theft accusations, open source is the perfect answer — “there is no IP ownership in an open-source world.” Recall the DeepSeek founder: his “entire emotional mindset was tied to open source.”
• And it’s working: DeepSeek is hosted by AWS and Google, successful in the enterprise, forked over 1,500 times on Hugging Face.
• The playbook Americans should recognize — open source as a defensive weapon: Google’s Kubernetes recruited the Linux Foundation and IBM against AWS’s lead, and “we don’t have a monopolist in that cloud business right now” perhaps because of that deft move; same with Android vs Apple and Meta’s Llama. It’s also pro-consumer — the closest thing to pure competition, “hyper competitive,” lowest prices.
• The wrinkle: Gurley hears DC may act to limit DeepSeek — both genuine security concern and self-interested lobbying. That opens a short US window for someone to “go left of DeepSeek in terms of openness” — Google or Meta plausibly; Anthropic has “been so anti-open source it would be very out of character.”

5. OpenAI goes open-weight; US compute is the goal

• Sam’s announcement: “a powerful new open-weight language model with reasoning in the coming months,” one of its first since GPT-2 — and asked whether they’d copy Meta’s 700M-user license restriction, he said no, a jab at Meta ahead of month-end LlamaCon. Brad hears Llama 4 lands this month: a 400B-parameter mixture-of-experts model (50-70 experts) with a ~10M context window.
• Gurley’s read (conjecture, he stresses — Brad’s the involved one): if OpenAI believes the battle is products over models, open-weighting is a deft move — it pressures rivals, drives pervasive global usage, offloads inference onto other people’s servers, and “wipes out the business opportunity for other models to play on the API side” — “the same reason Google would have supported Kubernetes.”
• Brad’s DC read: the administration “does not want to see a Huawei-DeepSeek Belt and Road either with chips or with open-source models” — they want the world running on US compute, US silicon, and Huawei 910 chips kept out. Gurley’s timeline: “the next 3 to 6 months could dictate who’s standing on top of the hill 5, 10 years from now” — and the executive shake-ups at Google, Meta and Apple “must represent anxiety.”

6. The $40B private IPO: 20x for the leader vs 50x Anthropic, 80x xAI

• Terms: SoftBank-led, ~$10B first tranche and ~$30B second, at ~$260B pre / ~$300B post; Altimeter participated in what Brad calls a “private IPO” bigger than almost any IPO ever done. His framing: “market leaders never look cheap” — he remembers Microsoft’s $15B Meta round being “roundly criticized as incredibly expensive.”
• The comps: ~$13B expected 2025 revenue implies a $15-18B exit run rate → ~20x forward run-rate. Anthropic at $60B on $1-2B revenue ≈ 50x; the X/xAI merger at ~$125B on ~$3B revenue ≈ 80x. The leader — which usually trades at a premium — is the cheapest of the three.
• Gurley’s needling: the nonprofit-to-for-profit conversion still has contingencies; headline numbers bigger than the piecemeal reality “smells of being promotional” — a message to Anthropic that “we’re here for the long haul.” From Uber-Lyft scar tissue (Masa at the table then too): “it is a sport of kings,” and unit economics get postponed — $5-7B in losses this year, roughly “$15 or 20 bucks a year for a non-paid user just to run the servers.” “If you’re going to lay chase to them, you got to be prepared to underwrite that cost yourself.”
• Brad’s rebuttal, worth keeping: in 2020 the headlines said Uber “would never be profitable… a failed business model” — it will do $6B in free cash flow this year. “The winner does take most.”

7. Demand is throttled and most users run a year-old model — the raise buys un-throttling

• The numbers: 1M ChatGPT users added in an hour, 20M paying subscribers, 500M weekly actives, 80-90% consumer share — and they’re rate-limiting demand (Sax: America’s AI leaders are all “being forced to impose rate limits”). Brad: two or three finished models plus agent products sit “on the shelf” for lack of GPUs; Sam’s “DM me if you have 100,000 GPUs” was real need — “the round was already raised.” Gurley: “It could be both.”
• The anime image-gen frenzy — a billion photos a day, alongside Gemini 2.5’s same-day launch — isn’t a bear case; Brad invokes Chris Dixon: “the next big thing will first appear as a toy.”
• On the ~4% paid conversion: more compute lifts everything, and the model mix converges — expect pricing tiers, eventually advertising (which Sam dislikes), and one smart model, “chat GPT5 or chat GPT6,” choosing for you instead of a menu of “04 mini and 03 and 01.” Brad’s pattern-match from early Google and Meta cohorts: this looks like a winner-take-all consumer application with network effects kicking in — and enterprise rides along via consumerization, the boardroom “get me an iPhone” moment replayed.

8. Agent economics break Google’s LTV math

• Gurley’s structural insight: Google is susceptible precisely because it maximized revenue per visitor. Booking.com will pay $50 against a $20 take on a $100 room using “LTV math” — repeat visits justify losing money upfront. In an agent world that dies: a white-label service underneath the model can share at most $10 of a $20 transactional take. “I just don’t think there’s any way you can get there” — “a huge disruptive advantage for OpenAI.”
• Brad completes the trade: $10 is all upside for OpenAI while it replaces $50 for Google (already down from $200 to $150). And the irony: “finally in 2025, the first time we’ve actually had competition for Google… now we get around to talking about breaking up their search monopoly. It’s ridiculous.”
• The standards angle: MCP — representing your service to a model rather than being scraped — is believed to have started at Anthropic, now backed by OpenAI; whoever’s most aggressive with open standards may get to define them. The through-line: tectonic plates shifting for the first time in 20 years of search dominance — “things happen very slowly… and then they happen very fast.”

9. CoreWeave flew into a category-5 hurricane — and the depreciation fight

• Broke issue Friday to ~$37 (“I was pretty damn nervous”), then a Google deal — Nvidia Grace Blackwells supplied through CoreWeave — took it to ~$60, answering the Microsoft-dependence critique: customers now span Microsoft, Meta, Google, OpenAI, Nvidia, Cohere, Mistral. Altimeter was one of the largest IPO buyers; credit to Mike (likely Intrator) and his team — zero to $30B+ in five years.
• Gurley on timing: anyone who says wait for the right market — “I would tell them to shut the [–] up and take your company out. You can’t control that thing.” All of 2024’s +30% sunshine produced no IPOs anyway. His riddle: Salesforce, Netflix, Square, Amazon, Palo Alto, Facebook, Snap, Proofpoint, NetSuite and CoreWeave all broke issue — judging quality by the IPO pop is “just such a wrong way to look at things.” (He also teases the Series C put at ~$38.75 and the day-one close at exactly $40: “engineered. Let’s be realistic.”)
• Brad’s GPU-longevity argument against Jensen’s “chief revenue destroyer” GTC comments (which implied maybe two-year useful life, not six): 2020’s A100s are still widely used in year five; OpenAI only just retired 2017’s V100s — a 7-year cycle — and CUDA updates make GPUs like his old Tesla, “it felt like a new car because my software got updated.” The contract math: 4-5 year guaranteed deals pay back capex, opex and GPUs in three years — year four is the margin, and consensus (~25% EBIT) credits nothing past contract end.
• Gurley’s pushback — worth keeping: “it’s not a zero or one… the earning power of that product drops over time,” so accelerated depreciation may better mirror useful life; incumbents stretched four years to six, and “the answer may lie somewhere in between.” Plus it’s a highly levered business that must delever.

10. TikTok NewCo: license the algorithm, unlock $600B of trapped venture value — if Xi allows it

• Brad’s sketch (he’s a ByteDance holder since 2015, not party to talks) ahead of the April 5 deadline: a “TikTok US” NewCo with ByteDance held under 20% (~19.5%), existing shareholders rolled in — 60% of ByteDance holders are US investors like Altimeter and likely Coatue — and ~50% new money from names like Amazon, likely Andreessen Horowitz and Oracle, logical since TikTok already runs on Oracle’s Texas cloud. Fresh capital goes into NewCo, which gets a license to the algorithm plus audit rights over algorithm and data; Trump wants the price low, possibly with a sovereign wealth fund piece.
• The likely bridge versus six months ago: founder Yiming had said the algorithm can’t be split — US creators feed the global product — so shutdown beat separation. Licensing one global algorithm to TikTok US may square that circle.
• The prize: ByteDance trades at ~$300B against fair value “closer to a trillion dollars,” certainly $800B (90% of value is Douyin, Doubao (likely) and non-US TikTok). Sixty percent of a trillion is $600B of locked-up US venture value — “more than almost every other unrealized venture gain put together” — convertible into hundreds of billions of DPI once the TikTok condition clears and ByteDance can go public.
• The catch, and the close: “I thought this deal would get approved by China. Now I’m not so sure” — Beijing could block it, and we just hit them with 54% tariffs. Bill’s optimistic note anyway: Trump wants a deal with Xi. Brad’s takeaway is that “US-China bilateral trade relations… will dictate the direction of global growth” for the next few years.