After listening to just two concalls, I feel far better informed, and in a much stronger position to make decisions, than after reading 10 AI-generated reports.

When Indian stocks are rallying, the sell side justifies valuations by running DCF models on earnings projected all the way out to FY50. But the moment something goes wrong, the same analysts slash near-term EPS estimates, switch to P/E or EV/EBITDA multiples, and suddenly discover X% downside.

Started travelling Delhi to Gurgaon 10 years back. Nothing changed and no hope anything would change here. I hope it hasn’t gotten worse now.

As I have been saying - In SMID, the low float is creating huge valuations problem. In bull market this is all what we want, but most of the companies making ATHs today will miss quarterly earnings - some due to their own poor management, some due to external factors. People are going crazy after one quarter's broad-based results. Last year's base was benign, inflation helped sell lower cost inventory at higher prices. Upper circuits are being hit some with minor good news and some without any news. Every stock is being justified by adjusting exit multiple higher. The wrong bags held today will revert to mean sooner or later, so selling at right time is very important.

Hopping on Namaste London Bridge listening concalls 🎶

Low float means valuations have no limit in India. Any PE can be justified. All one needs is a bull market, you buy at 50 PE, see it go to 100, and 100 PE becoming 150.

My focus has heavily shifted to global markets lately but thankfully @ishmohit1 @soicfinance has been making superb content to help out and set some baseline on recent things happening in India.

Very positive commentary from Talen Energy (operating in critical North Virginia hub of data centers) - "In late 2023 and early 2024 (when Talen contracted its Susquehanna nuclear plant capacity to AWS for its adjacent data center campus), fixed power contracts around $80 per megawatt-hour (MWh) represented a massive premium over prevailing wholesale grid rates, which historically hovered between $30 and $50/MWh. Now the forward wholesale prices for capacity" "Historically, power curves in competitive markets were “backwardated” (future prices dropped sharply because buyers refused to pay premiums far into the future). For a long time, traders in the forward markets (buying power for years like 2027 and 2028) were skeptical of the aggressive demand forecasts driven by AI and data centers. Now that real-world electricity usage is hitting record peaks in the cash (real-time) market, forward traders are finally pricing in that scarcity. As a result, the profit margins (spark spreads) for power plants in 2028 have jumped nearly 30%."

No single technology can defeat every drone; effective C-UAS requires a layered system. 1/ Drone defence is no longer just about shooting a target out of the sky. Threats range from £1,000 commercial quadcopters to autonomous military drones and coordinated swarms. Modern Counter-UAS therefore requires a layered “detect, disrupt and destroy” architecture. 🧵 2/ The first challenge is detection. Small drones fly low, move slowly and have a limited radar signature. They can easily be confused with birds or disappear into ground clutter. No single sensor is sufficient. 3/ A modern detection layer combines: • Radar to locate and track • RF sensors to detect control links • EO/IR cameras to confirm the target • Acoustic sensors to recognise rotor signatures The real advantage comes from fusing these inputs into one picture. 4/ Once identified, the preferred first response is usually a “soft kill”: • RF jamming breaks the pilot’s control link • GNSS disruption interferes with navigation • Cyber systems may exploit supported communication protocols and take control This avoids expensive ammunition and falling debris. 5/ But soft-kill systems have limits. They may struggle against drones that are: • Fully autonomous • Using frequency-hopping links • Navigating without GPS • Connected through fibre-optic cables • Operating as part of a coordinated swarm That creates the need for hard-kill systems. 6/ Hard-kill options include: • Nets and interceptor drones • Rapid-fire guns • Airburst ammunition • Surface-to-air missiles • Kamikaze interceptors The challenge is economic: using a million-dollar missile against a cheap drone is not a sustainable defence model. 7/ That cost mismatch is driving investment in directed-energy weapons. High-energy lasers concentrate light on a drone until its structure, sensors or propulsion system fails. They offer precision and a very low marginal cost per shot—but require clear line of sight and sufficient dwell time. 8/ High-power microwave systems attack the electronics instead of the airframe. Their key advantage is area effect: one pulse may disrupt several drones simultaneously. That makes HPM particularly attractive against swarms, although range, power requirements and friendly-electronics risk matter. 9/ AI ties the entire system together. It can fuse radar, RF, acoustic and thermal data to: • Classify the object • Assess its intent • Prioritise threats • Recommend the appropriate effector • Coordinate responses against multiple targets 10/ The future of Counter-UAS will not belong to one “silver bullet.” It will be a layered system combining: Detection → identification → electronic warfare → kinetic interceptors → lasers/microwaves → command software The winner will be the system that delivers the highest probability of defeat at the lowest cost per engagement.

Each GW of nuclear capacity requires 4000-5000 ton of piping spools and India is targeting to reach 100GW of nuclear by 2047.

Ambarella CEO says that coding is the leading usecase that drove the adoption of compute and led to build-up of data centers but no such leader usecase currently exists in edge-AI. He also says humanoid problem is harder than an autonomous driving car.

I think the main reason Google is not releasing models at same pace like Labs because for Google, the cost of failure is very high. They can’t release some untested stuff and wrap themselves in safety issues debate because of past regulatory oversight on its monopoly business model. Labs should take time to test themselves and not ask outside support for the same. If you are pioneering new technology or new product its your job to test for its safety.

At Amazon, if something bad happens and there is a customer impact, first thing that we do after mitigating the impact is doing a thorough root-cause-analysis RCA. We write a CoE (correction of error) document where we write the whole incident and write 5 why’s about why it happened. It is hard to escape this without doing properly as there are so many eyes on this doc. To mitigate the incident as quick as possible, we have proper logging, alarms, metrics to alert the owning team that something is wrong, so mitigation can be immediate. In some of the AI security breach incidents that have happened, it seems like models are getting released in hurry by frontier labs without significant guardrails, failure modes analysis. In this case, they should themselves slow down rather than asking outside intervention to ask everybody to slow down, until they fully understand the model behaviour.