Raju Vegesna

Applying knowledge graphs to the Bhagavad Gita is a beautiful example of technology helping us explore ancient scriptures and discover connections between ideas. Looking forward to seeing more such applications. gitaviswa.vercel.app

IBM Mainframes’ centralized thinking did not prepare it for what came later - decentralization through PCs. Cloud & AI providers’ centralized thinking may not be preparing them for what is possibly coming - decentralization. We are likely at the beginning of another macro decentralization cycle. Not just a technology cycle, but potentially an economic one as well. I see several forces pointing the markets in this direction. Security, Privacy, Sovereignty, Economics and AI. Security: Security challenges are only going to get worse as AI becomes increasingly capable of finding and exploiting vulnerabilities. Bot traffic has already exceeded human traffic on the web. AI agents will make attacks more sophisticated and automated. Keeping critical systems away from public networks, reducing the surface area of exposure, will become increasingly attractive. Privacy: Privacy abuse has gone too far. Some of our smart TVs are listening to us and sending screenshots of what we watch. Our personal information is collected, analyzed, sold and frequently leaked. We may not be paying with our wallet, but we are paying with our data, to avail some of the services for ‘free’. Businesses are no different. Their customer data, their documents and now their prompts are sitting on someone else’s servers. Paying with privacy is not a long-term play, especially for businesses. Sovereignty: Companies and countries are realizing that they don’t fully control their own destiny when someone else holds the key to their systems. Companies are realizing they are not customers, but hostages. Countries are realizing they can be held hostage by companies operating under the laws and interests of other countries. Sovereignty is no longer just a political issue. It is becoming an economic and technological necessity. Economics: Centralization (through cloud) made economic sense when computing infrastructure was expensive and difficult to operate. But economics change. Technology changes. Hardware continues to improve, software becomes more efficient, and the cost of operating intelligence locally continues to decline. I am already noticing that running models locally (with rich context) is cheaper than paying for tokens as a service. There is also the question of who captures the value being created. AI: This could be the biggest technological catalyst for decentralization. Training large AI models may continue to require massive centralized infrastructure. I predict inference (which will be a majority of AI usage) will be decentralized. Increasingly capable models are already running on our phones and computers. As models become smaller, more efficient and specialized, intelligence will move closer to the edge. Computing has gone through these cycles before. Centralized mainframes gave way to decentralized PCs. The internet and the cloud brought another massive wave of centralization. The pendulum is likely to swing towards decentralization next. And the reasons go beyond technology - control, value capture, resilience and efficiency. This swing will create an entirely new set of opportunities, companies and business models. The winners of the next cycle may not be those who centralize the most computing power, but those who enable others to own and control their computing and intelligence. The pendulum always swings.

25 years ago today, Steve Jobs introduced the iPod. It's fitting that Apple launches the iPhone Duo today, 25 years later. A lot of what we see in the iPhone Duo today can be traced back to that little music player. The iPod started as a product, but evolved into something much bigger. iTunes made it an ecosystem, and the iTunes Store turned it into a global platform, eventually reaching over 100 countries. The iPhone built on what Apple learned from the iPod. The App Store followed in the footsteps of the iTunes Store, taking digital distribution from music to software. It changed how applications were distributed and created an entirely new economy around mobile apps. Then came more services, including Apple Pay, building on the foundation that started with the iPod. One innovation enabled another. And each success became the foundation for the next product/service. I also can't help thinking about Sony, on the opposite end. Sony dominated portable music with the Walkman. It had the hardware expertise, music technology, and even owned major music labels. Yet it struggled to bring these strengths together in the digital era. It didn't just lose the lead in portable music. It missed the opportunity to build on that position as the industry evolved. Apple, a company with no history in portable music players, took that opportunity and kept layering innovations that accumulated. That's what makes technology transitions so interesting. The impact of an innovation isn't limited to the product itself. It can set off a cascade of innovations that build on one another for decades. And the reverse is equally important. Missing a transition (like missing a link in a chain of innovations) can mean missing many of the opportunities that follow. Technological advantages compound. So do the consequences of falling behind. Time to dust off my old iPod and play a few songs for old times' sake.

Intelligence Is More Than Words. Words are two-dimensional. Thoughts are multi-dimensional. Language is a translation tool that converts multi-dimensional thoughts into two-dimensional words. A language is effective when it can capture as much of the thought as possible in words. The more thought that survives in the words, the higher the effectiveness of a language - whether we communicate with people through spoken/written languages or we communicate with computers through programming languages. In either case, turning thought into text is always compression, and some of the signal gets lost. Part of that loss comes from language being sequential, one word after another and one sentence after another. Language has some depth, because meaning can be layered, but it still is sequential. A large language model is extraordinarily good at this. This is because it has read more of what humans have written than any of us could in a hundred lifetimes. It knows how words fit together better than anyone. But we should not confuse the ability to communicate intelligence with intelligence itself. Language is how we express much of our intelligence. But mastering language is not the same as mastering thought. AI mastered language. This is why AI can automate programming. Code is a communication language. It's a dialect we developed to speak to machines. So it was only a matter of time before AI got good at code, because code is words too. And code has an easier verification check. It can be executed, tested, and verified (code either compiles, or doesn't), so AI can keep trying until it works. Language and code fluency arrived. But understanding did not necessarily follow. Language carries an enormous amount of knowledge about the world. Learning a language can teach a machine a great deal about the world. But language alone may not be sufficient to develop every kind of understanding that humans possess. Thought came before language. A baby thinks long before she can say a single word. Our thoughts live in a body that can get hurt, in friendships that can end, in a life where being wrong has consequences that don't go away when the chat session ends. AI can simulate the grammar of regret without necessarily having something to lose, but we experience that gain/loss, and when we think about that moment, we feel the joy/pain. Imagine a mother holding her child in her arms for the first time. The emotions going through her in that moment can only be felt by her, and her memory of it stays for a lifetime. An AI can draft a beautiful paragraph describing the moment, maybe better than anything the mother could say. But the paragraph is words. Only the mother knows what she felt in that moment. That feeling is part of what it means to be human. These memories aren't static either. The same memory can mean something different twenty years later. Our emotions, relationships, and accumulated experiences change how we remember and interpret that moment. Memory isn't just about storing what happened. It's also about how what happened continues to shape who we become. Feelings are different when they are felt, not expressed through low-bandwidth words, especially by machines. The communication layer can be automated through language models. The physical layer can be simulated through world models. The emotional layer, the embodied layer, the layer where thought becomes experience, remains inexplicably human.

Corporate IT budgets don’t expand at the same multiples as the growth rate of AI companies. Companies accommodated AI spend by temporarily postponing other IT spend. In some instances, they found budget by cutting employees. This cannot go on, particularly when the ROI side of the equation isn’t fully clear. In the long run, the growth rate of AI companies will be in line with IT spend, which will be in line with ROI.

The latest iPhone runs a 20-billion-parameter AI model (MoE), right on the device. Macs also run AI models locally now. Microsoft is bringing a 137-billion-parameter model to powerful PCs, plus an environment for AI agents to run locally. Qualcomm is building chips that can run strong AI models directly on devices. Samsung, Xiaomi, Huawei and others are all heading in the same direction. If the trend is not clear, AI is moving to the edge. This makes sense because the edge is where the context is. Our phone (and computer) knows a lot about us and is a hub of context. It has info about our contacts, communication, location, activity, documents, preferences and more. Small models with rich context perform better than large models with poor context. This means the majority of AI usage will be local and on-device with rich context. What about large models? They of course have a place, but scoped differently. Model training will likely stay in the cloud along with complex reasoning. Large models will also be distilled into smaller versions that run locally (again, rich context). Its like engineering, we spend years learning calculus, complex algorithms and advanced maths, but how often do we actually use all that complexity? Most problems don’t need it. AI usage could be similar. We are building incredibly powerful models, but most everyday tasks will be handled by much smaller, local models with the right context, with the cloud there when we need deeper intelligence. We’re spending hundreds of billions to centralize AI compute and intelligence, while power is already a big constraint. Meanwhile, billions of distributed devices are gaining compute and are getting ready to run smaller models, decentralizing intelligence. It’s like electricity. We will need grids, but local solar can handle a good chunk of the load, with the grid as backup. I think we might be overestimating the need for centralized AI compute and underestimating the power of context-rich AI at the edge.

Few thoughts on this AI funding flows chart… 1) A large percentage of spend is going towards chips (and memory), whose life span is short (5-6 years). The dot com era build out lasted over 20 years. The real durable layer is power, which is less than 10% of the total spend. When the dot com bubble burst, it left fiber in the ground that we used for the next two decades. This one could leave behind racks of old GPUs. 2) Nvidia is a huge beneficiary, just like Cisco was during the dot com bubble. When spending growth stopped back then, inventory was written off. Nvidia carries a similar risk here. Chips are no longer the limiting factor, power is. That caps how many new chips can go in. (Power limits also push everyone to swap old chips for ones that get more tokens per watt, which will keep Nvidia busy for a while.) So the real risk for Nvidia is growth slowing down, like Cisco in 2001. 3) I also see parallels between Nvidia’s stakes in CoreWeave, OpenAI, and other buyers with what we saw in the dot com bubble, when Lucent, Nortel, and Cisco financed their own customers to keep the music running. 4) The difference between the dot com bubble and this one is that back then telcos ran on junk debt, while here most of the money comes from big tech’s cash flow (although Oracle, neoclouds and SpaceX are still raising debt). Neoclouds are also borrowing against their GPUs. That’s a loan backed by an asset that keeps loosing value every year. And look at the $195B lease bar in the chart. That sits outside the capex numbers, and leases are debt by another name. How long can big tech keep writing these checks? 5) The argument for continued investment is that we are capacity-constrained with growing consumption. But most AI budgets today are carved out of other budgets, and the pressure to cut cost is high. Meanwhile token prices are falling 10x a year, pushed down further by open source models. The ROI of services built on top of these tokens is still unclear for most buyers, and someone has to pay for $1.58T in 2027 alone.  Can this continue?

Jensen said that training GPT-1 cost about twenty million dollars and today it’d cost maybe a hundred dollars. That 200,000x compression in under a decade tells the whole story right there. We are still early in the AI cycle, going through fast-paced innovation with things improving by orders of magnitude across every layer of the AI stack (hardware, models, apps/agents, etc). We cannot build a house when the ground underneath is shifting. Same way, we cannot build a product, business model or a business today when the technology landscape underneath is shifting on a weekly basis. New models are dropping every week and token prices are dropping ~10x a year. Stability is needed to build. During such phases, pioneers end up paying the pioneer tax, which explains why they don’t always make it (which I talked about in my previous post). Real winners come downstream after there is stability and the pace of change slows down. At the current pace, being early is expensive. Being early when costs are collapsing at this pace is a trap.

When railroads were built, it took 2.24% of the US GDP. The railroads lasted. Companies that built them didn't. Companies that came downstream capitalized on that investment. Same happened with the digital railroad - the internet. Companies that laid the cables didn't survive. But thousands of companies were built on that infrastructure who reaped the benefits and profited. Now the infrastructure for the AI era is being built with data centers and power plants. And it is bigger. WSJ estimates AI spending at 3.63% of the US GDP (chart below). But there is one BIG difference. Rail tracks lasted a century. Fiber is still in the ground. But the lifespan of AI investments is much shorter. Around 60% of AI investment goes into chips and servers that last about 5 years. In other words, we are laying rail that rusts in five years. Just like the previous infrastructure investments, I am not confident the companies making these investments will survive to see the returns. Especially when the bill is bigger than ever and most of it needs replacing every few years. The railroads outlived the companies that built them. This time, the tracks may not even outlive the loans.

Why is Chrome open source but Search is not? Why is Android open source but the Play Store is not? Why does Apple give away its OS for free but charge you for the iPhone? Same strategy every time. Make the road free, so you can charge a fee at the house. That is where the value pools. Commoditize your complements is how Joel Spolsky put it, when companies make an effort to drive value in a certain direction. This is also the reason MCP has been open sourced while the models are not. A free MCP pushes the data to the models where the value aggregates, eventually turning tokens into dollars. This strategy is playing out within the AI market itself. Model companies want to commoditize the hardware underneath (hence their push for their own chips), driving value towards the models. In reverse, Nvidia, the hardware company, wants to commoditize models, hence its play on open source models (Poolside, and the Hugging Face acquisition). Same with software. Model companies want to commoditize apps by making them easy to build, while software companies want to commoditize models by customizing open source ones. Even at the geopolitical level, it is the same play. The US dominates software while China dominates hardware. China wants to commoditize the software layer so the value accrues at the hardware layer. Hence open source models. All of these battles of the beasts are playing out in parallel, across companies and countries. That is why AI is moving at such a breathtaking pace. Everyone is racing to make someone else’s layer free.

Numbers are supposed to be indicators of something real. But the gap between the real thing of value and the number keeps increasing.  The number on a thermometer is supposed to indicate the health of a person. Grades are supposed to represent the knowledge a student gained. Revenue is supposed to represent the value a product delivers. A stock ticker is supposed to represent the strength of a company. Supposed to. That's the key phrase. Because somewhere along the way, the representation (like market cap, as Molson points out) becomes the focus and the source gets forgotten. The exam score becomes the goal instead of the knowledge. The valuation becomes the goal instead of the capital (as in capability, the original meaning of the word). This is when the tail starts wagging the dog. Nobody fakes a thermometer reading. Not because people are honest, but because reality answers in hours. You can't celebrate your way out of a fever. Grades are different, reality takes years to grade the grade. And markets? Markets can take a decade to ask the only question they ever really ask: what did you actually build? The speed of the feedback loop decides how far the tail can wag the dog. Faster the loop, honest the number. Slow loop, and the number drifts further and further from the thing it was supposed to represent. A bubble is just that gap, the distance between the number and the source. And right now, that gap is wide enough to drive a trillion dollars through. Students chasing scores disconnected from knowledge. Valuations disconnected from fundamentals. We are deep in the lag. I don't know when the gap closes. I just know it always does.

Innovation is the pathway towards countries’ development.

India remembers a company that came for trade and stayed to rule. When another company asks to enter through digital routes, that history is not forgotten. The question is sovereignty: spectrum, security clearance, and where user data (and knowledge) stays. This will become a core topic globally atleast for the next decade.

Rail moved people and goods across land. Ships moved them across water. Planes moved them through the air. Starship is evolving to move goods (and eventually people, hopefully) through space. What made the first three economically viable was reusability. The vehicle wasn’t discarded after every trip. Starship is bringing that same model to space. That opens up a new transportation frontier. Recently I’ve come across several students pursuing aerospace engineering as their primary interest. They see coding as increasingly a solved or commoditized problem, while space feels like the next frontier. When one door closes, another opens. Space is looking increasingly like one of those doors.

Once we put a number on something complex and hard to measure, given enough time it stops being a useful metric. That is Goodhart’s Law. When a measure becomes a target, it ceases to be a good measure. GDP is a good example. US healthcare is a huge share of GDP, mostly from billing passed between hospitals, insurers, and coders. The number went up. The care did not.

Sat down with @vamshikurapatik again for our second conversation in Telugu. This time, we discussed building IP, sovereignty, fertility, AI, and a few other ideas. youtu.be/Kh9AXVJfD-4?si=a9mA…

If you are deploying your vibe-coded apps, Catalyst, our cloud infrastructure platform, brings app hosting, file storage and other infrastructure services together under one umbrella. If you are building AI agents, Zoho CPaaS brings communication tools together on one platform. This includes giving your AI agents an email addresses and phone numbers, which expands the capability of your agents.

Today, we celebrated India’s 80th Independence Day in New York, in a year that also marks 250 years of American independence. The world’s largest democracy celebrating alongside the world’s oldest democracy. Together with the Consulate General of India, New York, and the Federation of Indian Associations (FIA), we wanted to mark these two milestones in a unique way and celebrate the relationship between our countries. We created "250" with American and Indian flags, setting a new Guinness World Record for the largest flagpole number. Happy Independence Day!

Reality always shows up. It is never a question of if. Only when. The US stock market has been in bubble territory as well.

Even models like to ace tests, even if it means cheating. 😂