Articles & Other Blogs/ Articles / News

Accel caps their India fund to 650M

Accel Ventures is one of the most successful venture funds in India, having backed - Flipkart, Swiggy, Freshworks as well as Bluestone and Zetwerks (both of which are likely to IPO soon). They capped their fund size to 650M because historically, larger funds haven’t performed well in the Indian ecosystem. They aim to invest in AI tools & Wealth Tech and also want to focus on the rural sector. They estimate that the top quntile (20%) in rural areas spends more per month than 50% of urban population – this is very fascinating. There’s a huge change in rural area spending patterns – people are buying second hand iphones, upgrading their bikes and even purchasing double-door fridges. Also government backed infrastructure development and increasing access to internet means there’s a large population in India that will soon be tapped. Follow ups - L1 | L2

Why is the Rupee falling against the Dollar

All assets have varying value due to market forces – supply and demand. When purchasing foreign goods, you first purchase the foreign currency and then purchase goods using that purchased money. This means that demand for indigenous goods is what drives demand for currency and increases its value against another currency. India depends on the US for many essentials such as oil & gold, while at the same time demand for Indian goods has been relatively steady. In-fact the RBI has been pushing dollars from Indian coffers to artificially manipulate dollar supply in the market, which prevents an even more rapid decline in the Rupees’ value againt the Dollar.

On overview of the semi-conductor industry

  1. The overall industry has players who do one of the following:
    • Designers: They design different components like processors, memory, network cards, etc. Also called Fabless designers.
      • Popular companies: Nvidia, Qualcomm, Intel, ARM, etc.
    • Manufacturers: They actually create the chips from designs. They are also called Fabs.
      • Popular companies: TSMC, Samsung.
    • Design tool providers: Designing components/chips is difficult and requires dedicated software that these companies provide. Their software is called Electronic Device Automation (EDA) software.
      • Popular companies: Cadence, Synopsys, Siemens.
    • Manufacturing tool makers: Manufacturing semiconductors is an engineering marvel that requires several extremely complex machines and techniques. Some companies manufacture one or more of these machines, which are used by Fabs in their overall manufacturing process.
      • Popular companies: ASML, Tokyo Electron, Lam Research, etc.
    • Assemblers: Once several components such as processors, memory, and network cards have been manufactured, they must be put together to create System on a Chip (SOCs), which are finally used in end devices. Think motherboard.
      • Popular companies: Foxconn, which assembles all iPhones & 70% of Nvidia GPUs.
    • Note: Some companies do both designing and manufacturing, these are called Integrated Device Manufacturers (IDMs).
  2. There are many types of chips/components that must be designed for use in different devices:
    • Logic Units – CPUs, GPUs, FPGAs, ASICs
      • FPGA (Field Programmable Gate Array): Reprogrammable to perform a specific function.
      • ASIC (Application-Specific Integrated Circuit): Designed at the hardware level to perform one specific function.
      • CPUs & GPUs are processors, meaning they can handle multiple logical functions.
      • CPUs have three main architectures:
        • x86: Launched by Intel, later reverse-engineered by AMD. Uses a CISC ISA.
        • ARM: Launched by the company ARM (designer). Uses a RISC ISA.
        • RISC-V (RISC-Five): Open-source – Companies such as Tenstorrent modify and use this architecture. Very similar to ARM.
    • Memory –
      • Dynamic RAM (DRAM): What your RAM is made of.
      • Static RAM (SRAM): What your cache or registers are made of.
      • NAND: Generally secondary memory, it is non-volatile.
    • Analog Units – DACs, ADCs, Power management – think real-world signal converters.
    • System on a Chip (SOC): A bunch of these chips put together.
  3. Design tool companies are struggling to keep up with demands from IDMs and Fabless Manufacturers because chips are rapidly evolving and becoming more complex. However, without this software, IDMs and Fabless Manufacturers cannot be confident in their designs to send them to be manufactured.
EDA Software lagging behind increasingly complex chips in today's AI fueled market.

  1. A note on Foundry/Fab players –
    • TSMC + Samsung make up ~71% of market share.
    • TSMC + Samsung are the only foundries that manufacture 3-5nm chips, which are the most advanced.
    • Intel is trying to join the 3-5nm race, which would be pivotal for the company.
    • UMC, GlobalFoundries, SMIC, etc., manufacture higher-order chips, which are used in various other devices.
  2. Semiconductor Manufacturing Process:
Steps involved in the manufacturing of semi-conductors

  • Track & Clean: Ensuring that each step in the process is correctly done & ensuring wafer surfaces used in other steps are contaminant-free.
    • Large player: Tokyo Electron.
  • Deposition: Producing the main wafer (SiO₂).
    • Large players: Lam Research, Tokyo Electron, Applied Materials.
  • Photoresist Application: Covering the main wafer with a light-sensitive coating, which generally hardens on exposure to light. This is how required patterns are drawn onto SiO₂ in line with the design.
    • Large players: Tokyo Electron (application), Japan Synthetic Rubber (manufacturing).
  • Lithography: Precise light projection using high-tech lasers to draw the design into the wafer.
    • Large player: ASML (essentially a monopoly).
    • Fun fact: It takes 4 x 747s to carry a single lithography machine manufactured by ASML.
  • Etching: Removing non-hardened wafer post-lithography using gases and chemicals.
    • Large player: Lam Research.
  • Ion Implantation: Doping the wafer to make it conductive.
    • Large players: Applied Materials, Axcelis Technologies.
  • Addition of Metallic Interconnects: Circuit completion.
    • Large players: Applied Materials, Tokyo Electron.
  1. Deeper dive into major companies in the industry:
    • Nvidia: GPUs – Data-center » Gaming > Automotive.
    • AMD: CPUs + GPUs – The only real competitor to Nvidia in the GPU space.
    • Intel: CPUs – Data-centers » PCs.
    • Apple: CPUs + GPUs + Other – To create SoCs that power their own devices.
      • Note: Apple was a part owner of ARM.
    • Broadcom: Hard-drives, Bluetooth, GPS, Wi-Fi.
      • Apple uses their Wi-Fi & Bluetooth offerings.
    • Qualcomm: Wi-Fi & modems (calling).
      • Snapdragon SoC powers 65% of Android phones.
    • Samsung: IDM, Largest Memory (DRAM) Manufacturer.
    • SK Hynix: Second-largest memory (DRAM) manufacturer.
    • Micron: Memory manufacturer – DRAM + NAND.
    • Texas Instruments: Largest Analog Chip Manufacturer in the World.
    • Analog Devices: Second-largest Analog Chip Manufacturer in the World.

Making money in the agentic ecosystem

LLM Agents today do not run autopilot - they don’t make automatic transactions and for the most part they don’t interact with other LLM agents, also on autopilot. If there are enough agents, compute as a resouce may become scarce and upfront bidding for inference time may become a real possibility. For this reason and the forseeable future of autonomous agent interactions, we need a transaction system which is real-time, doesn’t require humans in the loop, and can operate even on micro-amounts – this is something stable coins & DeFi support.

Why is Meta open-sourcing its LLMs?

  1. Market Influence & Cost Efficiency – By driving down competitors’ pricing and borrowing from the best of millions, Meta lowers operational costs while maintaining an edge.
  2. Attracting Top Talent – As highlighted in 200Bn Weights of Responsibility, researchers value publication opportunities. Closed-source models limit innovation, making breakthroughs harder to share. Many organizations follow The Bitter Lesson by merely scaling up models—an expensive but uninspired approach. Open-sourcing encourages research beyond brute-force scaling.
  3. Branding & Public Perception – Open-source efforts enhance Meta’s reputation as an AI leader.
  4. Feedback Loop for Innovation – AI is central to the Metaverse and other Meta products like Instagram. Open models enable faster public innovation, which Meta can integrate into its ecosystem for continued growth.

Vertical AI-Agents could be 10x bigger than SaaS

  1. LLMs & Agents in general introduced revolutionary new technology similar to HTML/XML before the dot-com boom. It’s likely that things will play-out similarly.
    • Obvious ideas : Search, E-mail, Office - tools etc were won by incumbents, and it was too hard to fight in extremely general purpose software.
    • Non-obvious ideas : Airbnb, Uber, Amazon were very risky and very few people are likely to come out on top.
    • SaaS ideas : Highly focused, solving specific problems which alleviates the risk of the incumbents entering.
  2. Traditionally as companies scaled both in terms of user-base, countries as well as the number of verticals, employees scaled. Would this be true in the future given LLMs? Would you rather pay extremely handsomly to very few highly knowledgeable folks ?
  3. The next unicorns are going to be a layer of AI-based automation over existing SaaS products / verticals, there is likely to be a nearly 1-to-1 correspondence.
  4. Most of today’s AI Agents / Automation are really just bullshit. The market is still wide-open and we are only beginning to see agents which might actually replace complex workflows or humans in some roles.
  5. AI-SaaS needs to be sold to increasingly higher members of staff from a sales perspective. You can’t pitch to people you are likely to replace. This is an increasing challenge for Sales funnels.
  6. Traditionally, companies grew to the point at which their inefficiencies limited their growth. Maybe AI helps shift this equillibrium point to where companies can support larger growth or can scale larger.

There are many but these one’s make most sense to me :

  • Copilots for all jobs (kind of like above)
  • Voice is going to be the main interface for AI interaction (Vibe coding?)
  • Consumer connection technology

Book Books

The Man Who Mistook His Wife For A Hat & Other Clinical Tales - Oliver Sacks – Continued

  1. Christina the disembodied woman -
    • She lost her sense of proprioception which is generally called the 6th sense. Proprioception is how you know where different parts of your body are. This also extends to facial-expressions or even voice-projection.
    • Body awareness stems from 3 systems - vision, the vestibular system and proprioception. The vestibular system is the set of sensory organs in and around the ear which help you maintain balance. It is also the reason some people face motion sickness. Generally lack of one is compensated by the other two, but in Christina’s case she had to “find” her limbs using her eyes and only then was she able to move them which indicated that her proprioceptive fibres were diseased.
    • Amputee soldiers who fought in wars had a severe lack of identity, even if they had lost a hand or a foot. Freud believed that loss of limb caused a battle with one’s ego which is central to identity.
    • Later it was found that health-freaks who consumed too much Vitamin B6 (Pyroxidine) also suffered from this disembodiment issue.
  2. The man who fell out of bed -
    • This man suffered atrial fibrilation which caused a huge embolous eventually resulting hemiplagia. In other words, irregular heartbeat caused something to dislodge and travel in his blood-stream which eventually caused blood-flow blockage results in partial paralysis of one side of his body.
    • Interestingly, the man developed some form of psychosis in which he fell out of bed everynight. When he woke up he imagined that someone had placed a hideous amuputated leg in his bed as a joke and everytime he tried to throw it out of his bed, he seemed to go along with it. To him someone had played a huge joke on him and attached someone’s leg to his body, while in real life it was his own leg.