partner @wing_vc investing in AI applications, infra and physical AI. opinions, analysis, and banter on technology and business

New York
AlphaSense is post-training models end-to-end for search and research per SVP of Product @cackerso. The research loop is 80% of cost and 70% of quality in their agents, and AlphaSense expects its specialized research model can cut costs 40x.
2
5
1,197
Interesting work by Anthropic on exposure to digital and physical AI across occupation groups
3
19
1,517
PayChex CEO on agent adoption and sprawl: - They had 600 active AI agents a quarter ago and have 2,000+ now - 10K+ employees have adopted agents internally - Agentic payroll has cut manual processing by ~20%
7
3
24
3,279
Qualcomm announced a new chip Snapdragon 8 Elite Extreme Gen 6 (quite a name!). It can run a 30B-parameter MoE model fully on-device, the joint largest that anyone has announced so far.
1
2
8
2,434
Renen Hallak, CEO of @VAST_Data, on current state of AI storage demand: - An AI cloud said it needed 500 petabytes over three years. Revised it one quarter later to 2.5 exabytes - big labs moving into triple digit exabytes from tens of exabytes
1
7
24
3,297
Some good stats on today's private markets from Warren Gibbon: • Hyperscalers are ~70% of total US corporate capex this year • 12 firms account for ~3/4 of 2026 venture fundraising • Mid-market buyout multiples are 12x EBITDA vs 8x a few years ago (h/t: @davidweisburd podcast)
7
6
52
4,590
Goldman says that ~half of S&P 500 growth in EPS in 2026 is from AI investment. Hyperscaler Capex this year will be ~$800B (+94% y/y). Goldman expects growth going forward to be more muted at $1.2T in 2027 and $1.4T in 2028.
4
8
37
2,701
DeepSeek's revenue run rate hit $1B, more than doubling from <$500M in July. Cloud API gross margins are >80% despite the relatively low token prices they charge on their models. (via @theinformation)
4
5
46
5,876
TIL about "Dopamine Websites", which are websites that allow users to go through the online shopping experience without paying or actually receiving items. The trend started in Korea and has now expanded beyond it Some dopamine websites: - dopamine-shop.com - foodnevercomes.com/
4
11
2,413
Jefferies' satellite check on US data center buildout (via SynMax data): • ~11 GW in 2025 → 16-18 GW in 2026, 2027 capped in the low 20s, well short of 80GW of commitments • ~80% of the 2028 pipeline hasn't broken ground
6
19
138
18,204
CEO of Sequence Holdings @mjlee_2014 on the impact of their approach at BankSouth, their first major partnership/minority investment: - Consumer underwriting time down 94% since March - Average loan cycle went from 30 days to 11 days - Loan volumes handled by a smaller underwriting team despite growing 2x in Q2 v Q1
4
3
52
8,513
OpenRouter blocked 10x as much token-fraud dollar volume last month as the month before. What the fraud looks like: - stolen credit cards - reselling traffic against ToS - hacked and compromised accounts - reselling inference on the side
3
24
2,915
Clay CEO Kareem Amin on state of their business: - targeting $240M ARR this year, doubling next year - almost 18,000 customers - goal is to evolve product an agent that outputs "next best action to grow"
4
2
100
15,364
Pocket FM, an audio drama platform, is a fascinating example of how AI can transform content business. Content production is now 80x cheaper, and output is up from ~100K hrs/yr to 2.5M hrs/yr. As a result, retention is up from 44% to 76% and revenue runrate doubled to $500M
3
2
66
4,950
Affirm replaced ~15 years of tree-based underwriting models with a transformer-based architecture. It beat the next planned improvement in the prior architecture by 2x.
10
9
125
13,496
Costco CFO on AI search volume and intent: "Traffic to our site from AI search grew triple digits for the second consecutive quarter and continues to show the highest conversion rate of all site traffic"
12
23
266
33,365
.@gabepereyra of @harvey on when app-layer companies should train their own models: Use closed-source models until you hit the scale where you 'cannot economically serve intelligence at this scale.' Harvey's inference bill was in the hundreds of millions when they post-trained Tenet
5
2
31
3,554
China's 4 AI-chip "little dragons" are all public now. Market cap / LTM revenue: Enflame: ~$31B / ~$270M MetaX: ~$34B / ~$305M Moore Threads: ~$30B / ~$378M Biren: ~$12.6B / ~$330M
3
5
44
8,061
An investor day scheduled for the day after after you lost your CEO to Meta. Rough.
4
4
20
3,555