Lot of businesses with below mental work end up working best:
Supply a relatively small fraction of the end product's cost, yet purity at parts-per-billion levels, or tolerance levels for a deviation are low aka multi-step reliability is mission-critical.
Gemini 4 on Antigravity is different beast
🚨Gemini 4 Argon @GoogleDeepMind with @antigravity
has made a significant leap on our RSI-Inference Bench rsi-inference.com with a close tie with GPT-6-Astra.
It's the first model outside @AnthropicAI & @OpenAI to pass a Day-0 Model Support task ✨
Congrats @GoogleDeepMind @PichaiSunder 👏!
Simrat retweeted
🚨Gemini 4 Argon @GoogleDeepMind with @antigravity
has made a significant leap on our RSI-Inference Bench rsi-inference.com with a close tie with GPT-6-Astra.
It's the first model outside @AnthropicAI & @OpenAI to pass a Day-0 Model Support task ✨
Congrats @GoogleDeepMind @PichaiSunder 👏!
We just added Google's Gemini-4-Argon to RSI-Inference Bench 🎉
Lands at #6 of 14 on the joint leaderboard, right behind GPT-6-Astra.
•Day-0 support: 6th, 81% of hidden tests passed (Pass@1 6.7%)
•Inference optimization: 4th, 1.39× speed-up
Spotify is creating friction in ad supported tier in India to push users to become paid subs. Music to the ears of music labels in India 🎵🔥
Pratt (the aircraft engine maker) is integrating AI-powered engine inspection software to accelerate borescope analytics. Already in use on legacy programs, this capability is being implemented across our MRO networks and has reduced overall inspection time by more than 30% where it’s been deployed.
Diamonds and stones certification co in India is using AI to strengthen its moat, growth and margins:
Someone asked whether AI could lower barriers to entry for small competing labs.
1. Management’s answer highlighted data scale: machine learning in gemology requires millions of high-resolution stone images paired with multi-grader human verification.
2. Grading 4.35 million stones per quarter gives IGI a proprietary training dataset that no startup lab can replicate.
3. Operationally, they have deployed AI-based work allocation and scheduling across its laboratories, alongside ML-assisted grading. This keeps turnaround times tight (2 to 3 days in India even during peak festive rushes) and allows the existing team of gemologists to grade higher volumes without linear headcount growth.
One of the major Copper company listed in US which mines in Latin America, says that usually miners pay smelters to get more and more copper out of mined raw ore concentrate but now the smelters are paying miners as smelters are abundant and miners and specifically high quality copper ore capacities are limited.
This also brings need of players which are play on falling ore grades. Someone selling grinding balls has a structural tailwind to churn out more and more copper from low grade rock.
Its new mines also need new 220 kV lines, substations and water, which are anyway constrained due to DC buildout.
Scarce mine supply pushes buyers toward scrap.
Few interesting points came in Nokia concall which sits above optical fibre companies:
1. When GPUs are scarce, a better network lets customers get more use from the GPUs they already have. So network spending can increase even when GPU supply is tight.
2. Telcos are doing capex for long haul and subsea networks. These carry traffic between data centres on long-haul and subsea routes. So part of what Nokia calls “telco” demand is really AI demand.
3. Distributed inference is driving significant demand leading to more optical content. More smaller DCs will come for inference as large sites are less and less available and more smaller DCs make the network more resilient, secure and robust. Same point came in optical fibre company concall in India
AI architecture requires significantly more physical connectivity than traditional cloud computing. The number of fibers per switch rack has skyrocketed from roughly 500 in the traditional cloud era to 12,000 for today's 72-GPU AI nodes, and is heading toward 16,000 fibers in upcoming architectures. This scale-out, scale-across, and scale-up dynamic makes high-density optical fiber strictly non-negotiable.
Diamonds and stones certification co in India is using AI to strengthen its moat, growth and margins:
Someone asked whether AI could lower barriers to entry for small competing labs.
1. Management’s answer highlighted data scale: machine learning in gemology requires millions of high-resolution stone images paired with multi-grader human verification.
2. Grading 4.35 million stones per quarter gives IGI a proprietary training dataset that no startup lab can replicate.
3. Operationally, they have deployed AI-based work allocation and scheduling across its laboratories, alongside ML-assisted grading. This keeps turnaround times tight (2 to 3 days in India even during peak festive rushes) and allows the existing team of gemologists to grade higher volumes without linear headcount growth.
Simrat retweeted
Chamath reveals 1% of companies are 90% of AI token spend, and 80% of them are spending money from unprofitable VCs
"So the cycle here, is VCs raise from pension funds and LPs and foundations and family offices. They put money into a startup. That startup takes that money and invests in growth."
"Part of growing is by consuming other people's AIs. And so that company then starts to show revenue traction. Those folks then say, 'Hey, look how well we're doing,' raises a bunch of money from the same VCs. And it's just this circularity."
"Why that's important is what Ken Griffin is pointing out is that for AI to really scale as a predictable driver of long term earnings, it has to diffuse out of the startup ecosystem."
"You can't be in a situation where on the one hand, you're talking about breathtaking valuations, but 75 cents of every dollar is actually just a startup sending it back and forth so that they can mark themselves up. That's, I think, a structural problem."
"We've seen that structural problem in the past. That happened in the early days of cloud, when AWS initially started to expose EC2 and S3. The consumers of an EC2 instance or an S3 bucket were startups.”
“But the difference there is that AWS was abstracted from Amazon retail, which was already then a roughly break-even profitable business, at least on a free cash flow basis."
"So what's different here is you have a bunch of money-losing entities just spinning a bunch of money around and it has to diffuse more broadly."
Some good points coming on skewness and nature of current AI demand. Bubble or no bubble, one must keep eyes open.
Some very good commentary from Lam research on WFE demand, new proxies have emerged in India for this sector:
Wafer fab equipment (WFE) spend is about $155B in 2026. Lam expects more than $200B in 2027, or about 30% growth. Lam also expects to grow faster than WFE for the third year in a row.
AI uses more compute. Weekly token use rose about 250x in 18 months. Agentic AI needs more compute, memory, and storage. Lam’s rule of thumb: every $100B of AI capex creates about $10B of WFE. Lam’s share of that $10B goes up each year because chips get more complex.
Supply is holding the industry back. Demand is not the limit. Physical capacity is.
Clean rooms. A fab building takes years. In 2026, clean room space was the main limit. Customers got more output from their existing space and pulled projects forward. That is why Lam raised its 2026 WFE view four times, from the mid-$130Bs to the mid-$150Bs.
Component suppliers. This is a new limit. Some Lam suppliers see demand grow much faster than Lam’s own growth, mainly in advanced packaging parts. Accounts payable rose about $1.3B, which shows Lam is buying hard from suppliers.
Each quarter can only grow as fast as the slowest supplier.
If these numbers are true, whole DC value chain should fire.
Fun to co-host the GTC Berlin pregame show with @dee_bosa and nice of her to invite me on the postgame show.
Deepseek is generating $38 in inference revenue per GPU hour off-peak. That is $100b per gigawatt. 2x this at peak.
Seems important.
The stocks which have done nothing in last 6M, if they post good results, no bear market will stop them from being outperforming in near term.
The stocks which have done all in last 6M, if they post good or great results, the bear market will first unwind the euphoria before it decides whether to resume the outperformance again.
This is valuations, valuations are not mere P/E, EV/EBITDA.
Again terms and conditions apply.