Friday, September 18, 2026

Back to today

From X

@convequity

Convequity on X

I do feel like people have cognitive dissonance here: Jeff Dean isn't only about the model application layer — he is the one who masterminded TPU, GOOG's incredible global distributed systems, and making efficient hardware + software actually work. The story isn't simply that models don't matter anymore for GOOG, that GOOG is all in on just AI infra as a more neutral player at the model layer, avoiding heavy investment that could get quickly depreciated by Chinese open source models because infra is the better business. It is simply that GOOG's current model effort is turning south at an accelerated speed. And they know this area is strategically important — otherwise Pichai wouldn't have fought so hard to keep Jeff and others from leaving. You can also excuse GOOG's strategic failure — not foreseeing this year's huge demand, not reserving enough TPU in-house for both model serving and training, and instead selling the capacity to Anthropic — as a smart move to profit from infra and shield away from model competition. It is just that GOOG's internal divisions don't collaborate with each other: the infra team sold out TPU to Anthropic this year in a negotiation done last year, while the model serving and training team never saw the demand coming and never worked that closely with the infra team. By the sheer fact that there are still people sugar-coating this story, I feel less confident about GOOG — although I am still quite bullish on its TPU and infra business. BTW, Anthropic is looking for building its ASIC with $AVGO directly, and also building and running its infra as well. Already they have pivoted away from directly renting GPU from GCP to hosting its TPU and building its kernel to juice up perf by 50% to 2X, and they will continue this optimization route just happening slower than OpenAI does.
@convequity

Convequity on X

Jeff Dean’s departure is a seminal moment for $GOOG I’ve been watching this closely and the picture keeps getting clearer — and more concerning. Google is still treated as the safe consensus AI bet. Full-stack ecosystem, consumer reach, TPU muscle. But the competitive status of its frontier models looks increasingly dubious. The one-way talent flow tells the story. Google has seen massive outflows to OpenAI and Anthropic (reports put the numbers in the high hundreds to over a thousand in recent periods) while inbound from those labs remains tiny. This is not normal churn. It signals a structural talent problem that shows no sign of being fixed anytime soon. Jeff Dean leaving with Sanjay Ghemawat, Quoc Le and Oriol Vinyals, while Demis Hassabis steps back from day-to-day leadership, is the clearest signal yet that the research edge has peaked. Top talent does not want to waste cycles inside a large, conservative organization when the progress window is this narrow. We’ve flagged Google’s cultural and leadership issues in our recent rebalancing work. This just cross-validates it hard. What’s especially striking is the internal dynamic. All signs point to Koray Kavukcuoglu having been central to the environment that produced these outcomes — yet he has now risen to Chief AI Architect and the unchallenged operational AI leader, reporting directly to Pichai. The research org is fraying rapidly while the executive who oversaw much of that period moves up the ranks. That is a strange incentive structure. Large orgs struggle to deliver breakthrough AI. Talent density, bold vision and nimble execution matter more than scale. Meta rebuilt around an independent lab. SpaceX is doing the same with xAI. Google may eventually follow, but the current setup still seems unwilling to fully accept the reality. That said, the TPU and infra side continues to execute well with far less leakage. The competitive edge is shifting toward infrastructure, chips and consumer distribution — not frontier model leadership. This remains the biggest news of the week for a reason.

TSOH Weekly Roundup (09/18/2026)

TSOH Weekly Roundup covers organized retail crime, Home Depot’s centralized operating missteps under Bob Nardelli, and research updates on…

QNT: The 100× Question

Quantinuum's improving logical-qubit error correction, 1,504 successful-circuit benchmark score and paid deployments at RIKEN, BMW and…

Deep Dive: Grab Holdings ($GRAB)

Grab generates 53% of revenue from delivery, 36% from mobility and 10% from financial services, using its Southeast Asian marketplace to…

Copart 4Q26 Business Update

Copart's U.S. insurance volumes fell 7.5% in 4Q26, largely from one customer loss, while international insurance units rose 11.2% and…

EFXT · long

Enerflex LTD

Enerflex, a $3 billion natural gas infrastructure and equipment company, blends contracted energy-services cash flows with equipment…

A2Gold: Research Note

A2Gold's McIntosh drilling found 0.23 g/t gold over 65.5m, 370m southwest of the existing resource, and extended mineralization 86m…

Publications we follow

Every publication on this site, hand-picked.

108 publications

Newsletters 91

X accounts we follow

Every account on this site, hand-picked.

71 accounts

Investor letters we follow

78 firms

Sectors

Ordered by how often our editor reaches for them.