Gautam

Making sense of Supply Chain & Macro. ex-AWS/Amazon. Views are my own

[Virat Kohli @imVkohli]

The only time you truly fail, is when you decide to give up.

[Gautam @GautamSW7]

@craigweiss Please share his full feedback, I’m guessing Steve had stronger notes than “pretty dope” 😂

[Gautam @GautamSW7]

This article captured the current AI capex race pretty accurately, the funny part is nobody has to be irrational. Every next bid can make sense on its own, and somehow you still end up with a $700B bidding war 😂

[Venky Ganesan @venkyganesan]

A framework to make sense of the biggest investment wave of our lifetime https://t.co/kUyc5STGZm

[samiksha @heysamiksha]

We're launching the Muse / Instinct for enterprise, backed by a16z Speedrun.

We believe the next big shift in work isn't the model or the agent. It's the interface.

General-purpose AI assistants are supercharging how consumers get things done. Employees still spend 60% of their day doing work about work, a third of their time finding things, and switch between apps 1,200 times a day.

Contxt works where your team does. Text it in Slack or Teams, call it in Claude or ChatGPT, or forward it your emails.

It has a dynamic memory across your tools, history, and people, and can do work directly in your browser, just like your team does. You get the 100x productivity of a consumer AI assistant with the controls, audit trail, and shared knowledge a company needs.

We're a Princeton-educated team of ex-NASA engineers, repeat founders, and quants.

[Gautam @GautamSW7]

Found this interesting @Meta paper today on multimodal AI. Anecdotally, the finished apps from frontier labs already feel much better than the bare APIs. This reinforces that LLM is only one part of the product, the architecture around it matters just as much.

[Manohar Paluri @manohar_paluri]

MMI was born out of a simple question, what modality should models reply with. Turns out such a benchmark didn't exist so after a year's grind MMI is out. Congrats to the team!

Paper here: https://t.co/L15xUoSgwa

Data here: https://t.co/ZFkVmUvEzw

[Manohar Paluri @manohar_paluri]

MMI was born out of a simple question, what modality should models reply with. Turns out such a benchmark didn't exist so after a year's grind MMI is out. Congrats to the team!

Paper here: https://t.co/L15xUoSgwa

Data here: https://t.co/ZFkVmUvEzw

[Gautam @GautamSW7]

@Wai_Xea Haha exactly everytime AI solves one constraint, supply chain finds the next line item

[Gautam @GautamSW7]

The AI race is slowly becoming an infrastructure race. Models still matter, but you need the power, cooling, transformers, racks, land and grid capacity to actually scale them. OpenAI’s former data center head moving to Nvidia feels like another signal that the bottleneck is moving out of the lab!!

[Wall St Engine @wallstengine]

OPENAI’S FORMER DATA CENTER HEAD JOINS NVIDIA

Chris Malone, OpenAI’s former head of data centers, has joined $NVDA as VP of its DSX Platform, which helps customers design and build AI data centers around Nvidia’s infrastructure.

Malone left OpenAI in August after joining in early 2025 following the launch of Stargate. He previously worked on data center infrastructure at both Meta and Google.

[Wall St Engine @wallstengine]

OPENAI’S FORMER DATA CENTER HEAD JOINS NVIDIA

Chris Malone, OpenAI’s former head of data centers, has joined $NVDA as VP of its DSX Platform, which helps customers design and build AI data centers around Nvidia’s infrastructure.

Malone left OpenAI in August after joining in early 2025 following the launch of Stargate. He previously worked on data center infrastructure at both Meta and Google.

[Gautam @GautamSW7]

@alexandr_wang @clamasoto Mine is SAMI- Structured Artificial Multimodal Intelligence.

Structured because everything is organized, multimodal because muse handle text, images, voice notes and all 😊

[Alex Xu @alexxubyte]

Jev is TypeSafe AI's first System One Model.

It is 100x faster and cheaper than frontier LLMs.

That opens up a lot of use cases people usually skip because the big models are too slow or too expensive for them.

Here are 9 use cases where we think Jev could be used instead of an LLM.

[Gautam @GautamSW7]

Everything is a supply chain problem, AI included 😊 Alibaba wants 20GW+ of data centers by 2032. That’s power, land, cooling and transformers (the electrical kind) before the next wave of models can really scale. Time to watch the interconnection queue and not just the parameter count!!

[Wall St Engine @wallstengine]

$BABA UNVEILS NEW AI CHIP, PLANS 5T-10T PARAMETER MODELS AND 20GW+ OF DATA CENTER CAPACITY

Alibaba unveiled its new Zhenwu V900 AI chip, designed for both model training and inference, with performance roughly 3x higher than its predecessor released just four months ago.

The V900 is expected to enter mass production and commercial release in Q1 2027. Alibaba says clusters built around the chip can scale to as many as 500,000 accelerators for training and running frontier AI models.

Alibaba is also targeting more than 20GW of global data center capacity by 2032 and plans to launch its new Yitian 720 and 730 CPUs for agentic AI workloads in 2027.

Qwen4 is currently in training and expected soon, while future Qwen generations are expected to scale to roughly 5T-10T parameters, versus 2.4T for the current Qwen 3.8 Max.

Alibaba has committed more than $53B over three years toward AI infrastructure and capabilities.

[Wall St Engine @wallstengine]

$BABA UNVEILS NEW AI CHIP, PLANS 5T-10T PARAMETER MODELS AND 20GW+ OF DATA CENTER CAPACITY

Alibaba unveiled its new Zhenwu V900 AI chip, designed for both model training and inference, with performance roughly 3x higher than its predecessor released just four months ago.

The V900 is expected to enter mass production and commercial release in Q1 2027. Alibaba says clusters built around the chip can scale to as many as 500,000 accelerators for training and running frontier AI models.

Alibaba is also targeting more than 20GW of global data center capacity by 2032 and plans to launch its new Yitian 720 and 730 CPUs for agentic AI workloads in 2027.

Qwen4 is currently in training and expected soon, while future Qwen generations are expected to scale to roughly 5T-10T parameters, versus 2.4T for the current Qwen 3.8 Max.

Alibaba has committed more than $53B over three years toward AI infrastructure and capabilities.

[Gautam @GautamSW7]

I think this misses the conglomerate model a bit. Tata has cash-generating businesses funding bigger bets, much like Amazon did before AWS became such a massive payoff. Not every bet has to work, a few just have to work really big. And I think the Tatas have managed sunk cost fallacy pretty well over time. If history has taught us one thing, never bet against the @TataCompanies 😊

[Gaurav Dalmia @gdalmiathinks]

To me as a businessman, this is the number one challenge at Tata Sons.

[Gautam @GautamSW7]

Amazon did ~$69B in ad revenue last year vs ~$269B from online stores. That’s the tension here, true agentic commerce makes shopping easier, but it can also bypass the sponsored discovery layer Amazon monetizes so well. A truly Day 1, customer-obsessed Amazon would probably be willing to cannibalize some of that ad revenue if it created a much better buying experience. The flywheel still turns, the revenue just gets redistributed.

[Craig Weiss @craigweiss]

amazon cutting off muse is a sign of an aging business

[Gautam @GautamSW7]

@myveroai Yeah in a way it does come back to traditional automation right?? may be the difference now is just knowing where you actually need a model and where simple logic is enough!!

[Gautam @GautamSW7]

This makes Jev click for me now. It’s not really about replacing frontier models, it’s about not needing one for every decision an agent makes. Let Jev handle the fast, repetitive decisions and bring in the frontier model when deeper reasoning or generation is actually needed. Maybe the agent stack ends up being less about one model doing everything and more about using the right intelligence for the right decision.

[Sarah Drasner @sarah_edo]

🎇 I made a small Chrome extension: a side panel that drives any site's WebMCP tools with Jev!

When you type, on every keystroke it picks the relevant page's tool, fills in the arguments, and tells you how sure it is.

Here it is grocery shopping:

[Gautam @GautamSW7]

@sarah_edo @OnFinality Got it, that makes sense. So the every keystroke behavior is more about exposing how Jev is evaluating in real time and not necessarily how you’d wire it in production, that’s helpful clarification, thanks @sarah_edo

[Gautam @GautamSW7]

Diesel at $6.28/gal is where Hormuz stops being just a geopolitical headline. Higher fuel costs flow straight into freight, landed cost and inflation. What interests me from a supply chain standpoint is the shift from managing disruption around a chokepoint to actually designing networks to depend on it less @EIAgov

[unusual_whales @unusual_whales]

JUST IN: The Trump administration ‌has proposed investing $5 billion in a new fund that would help Middle ⁠East countries rebuild energy infrastructure battered in the Iran war and reduce their reliance on the Strait of Hormuz to ‌transport ⁠oil and gas, per WSJ