Friday Chatter (20250425)
Friday Chatter features anonymous conversations between two or three people discussing industry rumors and providing both forward- and backward-looking insights into the market.
Google
Q: Google reports earnings this week, and it feels like the numbers are already baked into expectations. What’s your take?
R: I’d call it a perfectly ordinary print—nothing spectacular. Expectations were bleak: everyone was down on Search; YouTube was expected to miss; GCP wasn’t supposed to shine. So the “beat” was mostly a function of a low bar. On the call, every analyst—each in their own wording—asked whether Google’s CapEx would change if the macro backdrop deteriorates. Pichai and the CFO repeatedly made it crystal-clear: CapEx stays. If the economy weakens, they’ll squeeze headcount and comp—the same layoff-plus-efficiency theme we’ve seen across tech for two years.
Q: Makes sense. They’re the biggest company facing the Innovator’s Dilemma in our lifetime. Yet they’re well positioned: a cash printer, ample corporate fat, and two legendary founders people forget are still top-10 billionaires. Watching the corporate immune system fight the dilemma is fascinating. Gemini 2.5 is solid, but look at Claude’s “Artifacts” (OpenAI calls it Canvas): a pane to stash generated code or content. Super handy, yet neither the web nor the mobile app has it—no code highlighting, nothing. A perfect mirror of their inertia.
R: They also have some—let’s say—“creative” AI metrics. Sundar announced that AI Overviews are wildly successful, rolling out to all of Search and monetising roughly on par with legacy results.
Q: They have to claim that—it’s the keystone of their Innovator’s-Dilemma survival plan. I always click the citations because I know LLMs hallucinate, but an unsophisticated user might not. AI Preview itself is a destination: it doesn’t push you elsewhere, so no advertiser is bidding for that click-through. Old-school Google tried to eject you from the site; AI Preview wants to keep you there. That’s the dilemma—no sugar-coating it. Internally and externally they must insist it works; otherwise the plan collapses. We outside the PR bubble know it will ding margins, but as long-term investors we also know it’s the only path they’ve got.
R: Exactly. GCP still trails AWS and Azure. AI helps, but the 2020 “cloud optimisation” mantra is coming back. Add accelerated depreciation: CapEx was $50 bn in 2024, guidance is $75 bn this year—over 30 % growth. Free cash flow will bleed; there’s no dressing that up.
Q: Early last year people still scrounged for A100 80 GBs. By January it was obvious: outside tiny training clusters, nobody uses A100 for inference. They keep them for training because the clusters exist and A100 depreciation is gentler than H100’s. H100 doubles performance but rents at $1.50/hour; A100 is $1.00, so no one chooses A100 for new work—even on a two-year horizon.
R: Right—the A100 shipped in 2020, so we’re three years in. Call it done.
Q: H100’s trajectory is different, especially once B200 arrives. There are more H100s than A100s in the wild; two years from now, with B200 mainstream, H100 rental could fall below $0.50/hour. That’s a brutal curve—nothing this steep since the Intel CPU refresh frenzy a decade ago. GPUs went from a four-year life to something far shorter.
R: Quantity matters too. CSPs still depreciate GPUs over five-plus years on paper. Anyone running the racks knows the real economic life is maybe three. That’s right in Jensen’s sweet spot. Cloud providers are squeezed unless they can pass the cost on. Rumour is Google is leasing CoreWeave’s H100 fleet.
Q: Makes sense.
R: Word is the rate is under $1.50 per GPU-hour—basement pricing.
Q: Yeah, bizarre: H100 at $1.50 while a 5090 gaming card rents just above $1. Pricing is all over the place.
R: Only way CoreWeave can go that low for three-plus years is if those servers are fully written off—no financing cost, just power. Then sure, rent them; otherwise the math collapses.
Q: Fair point. If the boxes are off the books and Google signs a long contract, everyone wins. Anyway, shifting gears: Google’s ad revenue shouldn’t be affected by the May 1 “Liberation Day” timing in Q1, right?
Walmart
R: Let’s be clear: Google never gives formal guidance. What we got was a Q1 snapshot. Analysts tried to pry into how ads are trending so far in Q2, but management stone-walled: “We’ve only seen a few days—too early to comment.” In short, don’t read anything about “post-Liberation Day” demand into this print.
Q: Then what about Facebook—Meta? What’s the setup?
R: Meta’s chart is wild. From the YTD peak around $740 it cratered to $500 after Liberation Day—-33 %. It’s crawled back to ~$550, still a hefty discount. If a recession really is coming, that pricing looks rational: ad revenue is the first casualty.
Q: But are we even heading into a recession? Walmart and Target just told suppliers to keep shipping.
R: Right, huge signal. My read: there’s >50 % chance the retailers have an implicit promise from D.C.—either tariffs get paused or resolved before those China containers hit port. Yes, Walmart tells suppliers, “We’ll cover tariffs,” but they still fear demand destruction if price tags jump. No one wants a 2022-style Target inventory write-down. More plausible: the White House quietly assures them “Give us two-three weeks; we’ll sort it.” Coincidentally, the CEOs met with Trump last week. Net-net, odds of a U.S. recession just dropped.
Q: They could just get a carve-out—retailer-specific exemptions.
R: Politically radioactive. It would look like Washington hand-picking winners and losers. Every food-and-beverage importer still paying tariffs would howl. Even GOP rank-and-file would balk.
Q: Politics aside, if tariffs stick, Meta eats it more than Amazon. Amazon’s alphabet-soup private-label sellers just mark the import at $1 and pay $1 duty—no big deal. The hit lands on branded SMBs—the very DTC players who scaled on heavy Meta/Instagram spend. Think fashion upstarts, Anker/Roborock-type electronics, Narwal vacuums. Their ad budgets vanish first, so Meta hurts disproportionately.
R: Fair—Meta would be the biggest casualty. Still, I think targeted exemptions would be a political nightmare.
Q: This administration’s racked up “disasters” all year and keeps polling fine as long as deportations and D.O.G.E. cuts. I doubt SMB pain moves the needle.
R: Deportations don’t spike unemployment. Smacking SMBs might—and then unemployment headlines get ugly.
Q: Which would shove the Fed toward cuts: Walmart/Target hold prices, inflation stays tame, unemployment ticks up—perfect cover to ease.
R: We’ll agree to disagree. I can already picture cable news parading irate shop owners 24/7—“government picked winners!” GOP would bleed, Trump popularity or not.
Q: Realistically, no way we ink a China deal in three weeks.
R: Maybe just a tariff pause.
Q: A deal with India, sure. China—maybe.
R: Let’s call it “maybe.”
Q: Total coin-flip. We’ll see.
R: Speculation aside, “keep shipping” is the only hard data. With that, recession odds look meaningfully lower than two weeks ago. Even if every China-linked DTC brand quits Meta overnight, that’s maybe a 3–4% revenue hit—and auctions re-fill empty slots. Worst-case is 4%, which is extreme. So Meta isn’t as doomed as the tape implies. My bigger worry is their AI strategy…
Meta
Q: LlamaCon is next Tuesday.
R: Next Tuesday—got it.
Q: Yup.
R: Then we’ll finally see their next-gen “reasoning” model.
Q: Speakers are Satya and—pretty sure—the Databricks CEO, plus Meta folks (Chris Cox, Manohar, Angela Fan). So: new reasoning model, plus some shiny consumer features—likely baked into the main apps. We’ve already had “Llama Maverick” and “Llama 4 Scout.” Maybe Behemoth lands, maybe it slips. Rumor mill says Qwen 3 and DeepSeek R2 could drop Monday on the China side, so we’ll see a flurry.
R: DeepSeek already pushed an R1T—basically a bargain-bin R1 model.
Q: End of the day, whatever DeepSeek ships next won’t leapfrog. Even with 996 schedules, you still need bake time. GPUs grind 24/7, but you have to nail data mix, curriculum, RL-fine-tune, etc. Meta’s habit of setting a hard LlamaCon date before research is done makes me nervous. In contrast, Google doesn’t pin model launches strictly to I/O Day. Credit where due: FAIR still publishes gems—yesterday’s paper used an LLM as a black-box optimizer to iteratively guess what’s in a picture from abstract feedback. Super-clever. Meta has deep research; it’s just siloed. ByteDance? No research culture—seven or eight competing LLM teams throwing spaghetti at the wall. Alibaba similar. U.S. players run fewer parallel bets—Google has two-ish big-model teams and, hey, Gemini/Gemma 3 turned out great; lots of shops already replaced Llama with Gemma 3. Google stays research-driven; product follows. Anyway, I’m rambling.
R: Fair point: prices shape stories, not vice-versa. When Meta ripped 20 days straight, the narrative was: “Meta’s crushing AI—models commoditize; he who owns distribution wins.”
Q: That thesis still holds.
R: Exactly. Same fundamentals, but a 30% drop and suddenly “Llama’s dead, Meta’s losing the AI war.”
Q: Distribution still trumps models. Look—OpenAI offered $3B for Windsurf after missing Cursor at $10B. All about user reach.
R: They also waved a hand: “Hey, if Chrome’s up for grabs, call us.”
Q: Political theater. If anything, it strengthens the case for Chrome staying inside Google. Browsers should move cautiously—no overnight LLM injections. I’m still salty Chrome killed JPEG-XL, but I’m a Firefox user so…
R: Meta still has the TikTok cloud. Honestly, TikTok will just keep operating until the next administration—no ban, no forced sale, no clean resolution.
Q: True, but Beijing explicitly said TikTok isn’t on the tariff table—you can tweak tariffs anytime; TikTok can only be sold once. TikTok keeps calling DC’s bluff, hinting at self-shutdown. Even when it disappeared from app stores for a month, the devs kept cranking. Uncertainty is toxic, yet China won’t swap TikTok for lighter tariffs—full stop.
R: Exactly—let TikTok slide and precedent says SCOTUS rulings are optional. Slippery slope.
Q: It’s the selective-enforcement circus: Congress said “delist TikTok,” but DOJ says, “delist and we sue you.” So Apple and Google keep TikTok and are technically breaking the law daily. Next president probably won’t litigate either. No private right of action, so the law is toothless. Total mess.
R: Exactly—dangerous slope all around.