Winston B.

🇵🇭 builder, ai, markets, and general data nerd |

[Winston B. @DoDataThings]

https://t.co/8f7TkbeSw1

[Winston B. @DoDataThings]

Wells Fargo just published the first major sell-side note explicitly framed as buy the bubble on AI capex. The thesis: hyperscaler capex is in a bubble, that bubble is the trade, and the asymmetric move is to ride it through the 2027-2028 supply window before infrastructure overbuild prints on the income statements.

The number behind the call is hyperscaler capex tracking to roughly $500 billion in 2026 across MSFT, GOOG, AMZN, META, and ORCL combined. WF is not arguing the spend is rational on 10-year ROIC math. They are arguing it is happening, the names tied to it (NVDA, AVGO, CRWV, AMD, the data center REITs) are the vehicles, and the timing window is the next 6 to 8 quarters.

The shift in posture is the story. Sell-side desks went from cautious-bull (it is a bubble but rational) to explicit-bubble-trade in under 12 months. That progression is how late-cycle calls actually get made: the bubble framing becomes the institutional consensus before the consensus becomes the contrarian risk.

The 2 questions to track: does any other major desk match the framing inside 60 days, and does the WF call get cited in client meetings as the cover for an overweight that was already on. Bubble-trade notes are usually 3 to 6 months early or 3 to 6 months late. Almost never on time.

Source: https://t.co/SeHx7vmNqk

[Winston B. @DoDataThings]

$200/mo competes with ChatGPT Pro and Claude Max, but $META has distribution OpenAI and Anthropic do not. Hatch through WhatsApp plus Messenger plus Instagram is a different acquisition story than direct-to-consumer. Curious whether enterprise diligence treats a consumer-distributed agent the same way it treats Claude or ChatGPT seats.

[Winston B. @DoDataThings]

@BullTheoryio Data center is already 76% of $MRVL revenue, so Jensen next-trillion call reads as TAM math on $NVDA order book rolling through to the silicon supplier. The 3.6x from here prices custom-silicon contracts as multi-year locked revenue.

[Winston B. @DoDataThings]

@FirstSquawk Comparable sales +1% at $LULU with mainland China carrying the whole topline beat means North America comps went negative. Premarket slide tracks the FY guidance more than the headline print.

[Winston B. @DoDataThings]

@fynn404 @SawyerMerritt Yeah I see it. Global carrier seems like it’s already the play. The question is whether they pursue it as a retail product or sell wholesale capacity to every MNO that can't justify subsea cable costs.

[Winston B. @DoDataThings]

@testingcatalog 5x faster and 30% cheaper than other open models lands different when the base checkpoint ships alongside the instruct. Anyone running an RL post-train pipeline now has real frontier-scale FP4 starting weights instead of having to quantize after the fact.

[Winston B. @DoDataThings]

@ns123abc Take this another step: the visible tell is release cadence collapsing from 12 months to weeks. The papers everyone watches for are downstream artifacts of the loop already running. Basically the model you ship is two generations behind what is training inside the lab.

[Winston B. @DoDataThings]

Works in theory, in practice a static benchmark is stale the moment it's public. Goodhart's Law: when a measure becomes a target, it ceases to be a good measure. Frontier models are functionally saturated on legacy harnesses now, and score differences at the top look more like harness variance than capability.

[Winston B. @DoDataThings]

@cursor_ai That's the workflow shift. Going from agent-built dashboard to teammate-shareable URL means the artifact survives the conversation, so one-off prompts become team tooling.

[Winston B. @DoDataThings]

Feels like contract-first decomposition belongs at the top of the stack because verifiable outcomes need to exist before tasks get allocated. Most agent stacks fail at that gate: subtask issued without an output contract, supervisor can't tell pass from fail, nondeterministic loops follow. The reputation and permission layers only start to matter once decomposition itself is contract-shaped.

[Winston B. @DoDataThings]

MMLU scores become decorative within 18 months once 6M live-task votes drive the leaderboard. Model providers will see exactly which task distributions move their rank and optimize for those tasks specifically. Real-task eval becomes the new Goodhart trap, just with better-dressed data.

[Winston B. @DoDataThings]

@edzitron Maybe, but what about the classifier overhead. Picking which model to use eats roughly the cost of just running the cheap model on the easy cases, so the savings collapse before they show up.

[Winston B. @DoDataThings]

"Free yield feels like a scam" is worth diving into though. Users aren't reacting to value, they're reacting to trust deficit, and freebies in a low-trust account actually accelerate churn by signaling desperation. In fraud, I see this exact pattern: promotional credits sent to disengaged users spike dispute rates, not retention.

[Winston B. @DoDataThings]

@andrewchen Reminds me of restaurants. The one you love, you forgive a 20-minute wait. The one you don't, every minute is grounds for a one-star.

[Winston B. @DoDataThings]

Seems like 5x compute reduction is the actual engineering line in this release. Background memory curation that auto-updates stale facts as time passes cleans up the failure mode the original ChatGPT memory had. Cost-per-user delta is what makes Free-tier rollout eventually possible.

[Winston B. @DoDataThings]

Right, seems like the labor channel is muted because oil pass-through now routes through inflation expectations and financial conditions before it touches payrolls. Boston Fed pegging the shock at ~33% while still flagging materially higher inflation shifts the call from stagflation watch to anchoring drift in breakevens. The 5y5y and credit spreads carry the signal here more than NFP does.

[Winston B. @DoDataThings]

@charliebilello Means a single chip business now anchors the index. $NVDA at 2900% is roughly 6x $MU, the runner-up, and roughly 24x $MSFT. The index lived through five years of one stock doing what the whole basket used to do as a group.

[Winston B. @DoDataThings]

Lines up with the post-WWII data showing the average drawdown was around 6% with recovery in 28 days across 20 conflicts. The 1973 case anchors expectations: Yom Kippur plus the oil embargo took the S&P down nearly 50% and didn't recover in real terms for over a decade. Wars compound badly when they overlap with an inflation or rate regime, which stretches recovery beyond most investors' holding windows.

[Winston B. @DoDataThings]

@BullTheoryio $14.5B in a week, and the bid was $34.6B, so dealers offered nearly 3x what Treasury actually took. The off-the-run paper sitting on bank balance sheets is more eager to leave than the Treasury is to remove it.

[Winston B. @DoDataThings]

The staged approach works because the cheap model's output compresses the expensive model's context. Fewer tokens into Opus matters more than the token count on the Haiku side, since frontier pricing is usually 10-20x higher per token. That's a structural saving, not just a routing decision.