To The Moon

I want to write this week about a topic that’s been bugging me recently. It’s an irrational (in my opinion…) behavior that’s starting to emerge in earlier stage startup land more and more. The challenge is - I can see how we are getting to an irrational end state with a rational starting point… And it all comes down to how employees view and value startup equity. So this post is meant for all of the employees evaluating different offers from startups. Let me set the stage with a hypothetical conversationFounder: “I need to raise at the highest valuation possible. And after I do, I need to raise again in 6 months at an even higher valuation.”Me: “Why?”Founder: “Because that’s how I will attract and hire the best talent”Me: “Why? Wouldn’t new hires want a lower valuation for more upside?”Fou

Over the last year we have all been trained (at least I have…) to look at model launches through one lens. What is the benchmark score and who sits at the top of the leaderboard. It felt like every product announcement was immediately reduced to a scatter plot and a few social media hot takes about who beat whom by a few points on MMLU or GPQA. But something subtle started to shift with the Gemini 2.0 family and became much clearer with this week’s Gemini 3.0 launch. The story is no longer only about raw model quality. The frontier is getting crowded, the performance gaps are narrowing, and the real competition is moving up the stack into platforms. Said another way, the next decade in AI will be defined not only by model breakthroughs, but also by distribution, integrations, and the shape

The AI FactoryIf I zoom out for a moment and look at the current trajectory of AI infrastructure, it’s hard not to see an entirely new pattern forming. Inference keeps getting faster. Inference engines keep getting smarter. And the ecosystem around them keeps getting more modular and open. What once felt like specialized machinery locked inside a handful of labs is now drifting into the hands of every company with a GPU budget and a few strong engineers.Neoclouds like $CoreWeave, Inc.(CRWV)$ and $TOGETHER PHARMA LTD.(TGPHF)$ have rewritten the economics of GPU access. Inference clouds like Fireworks, Baseten and fal have done the same for reliable serving (and we’ve already separated into separate infere

If I had to simplify software market cycles, I’d say they come in two phases: the expansionary phase and the consolidation phase.In the expansionary phase, buyers scoop up software almost indiscriminately. There’s little concern for cost or efficiency, what matters is speed. It’s about accelerating product development, capturing market share, or outspending competitors to stay ahead, all under the assumption that growth will take care of everything else. During this phase, public markets shift their focus entirely to growth over profits. Take a look at the multiples chart I post later on breaking out multiples by high, medium, and low-growth companies. You can see the high-growth bucket has seen multiple expansion this year, while the mid-growth bucket has seen steady contraction.In the co

This week the 3 hyperscalers reported ( $Amazon.com(AMZN)$ AWS, $Microsoft(MSFT)$ Azure and $Alphabet(GOOG)$ $Alphabet(GOOGL)$ Google Cloud). What did we learn? Most importantly - they ALL called out still being meaningfully capacity constrained. CapEx guides are going up, data center builds are going up, power constraints are meaningful. This isn’t the telecom bust where the world laid fiber that was “dark” (ie unused). GPUs are being used the second the come online…Here are the numbers:AWS (Amazon): $132B run rate growing 20% YoY (last Q grew 17%)Azure (Microsoft): ~$93B run rate (estimate) growing 39% YoY (last Q gre

Cloud Giants Update:AWS ( $Amazon.com(AMZN)$ ): $132B run rate growing 20% YoY (last Q grew 17%)Azure ( $Microsoft(MSFT)$ ): ~$93B run rate (estimate) growing 39% YoY (last Q grew 39%)Google Cloud ( $Alphabet(GOOG)$ $Alphabet(GOOGL)$ includes GSuite): $61B run rate growing 34% YoY (last Q grew 32%, neither are cc)ImageFor SG users only, a tool to boost your purchasing power and trading ideas with a Cash Boost Account!Welcome to open a CBA today and enjoy access to a trading limit of up to SGD 20,000 with upcoming 0-commission, unlimited trading on SG, HK, and US stocks, as well as ETFs. Find out more here.Other helpful

1. $Microsoft(MSFT)$ Azure at a ~$93B run rate growing 39% constant currencyQuarterly YoY growth trends below Line chart titled Azure YoY Growth (cc) displays blue line plotting quarterly year-over-year growth percentages from about 27 percent in Q1 2021 rising to 39 percent in Q3 2025 across x-axis quarters labeled Q1 to Q3 from 2021 to 2025 and y-axis from 0 to 60 percent with data points at 27 percent, 28 percent, 31 percent, 34 percent, 39 percent Estimates of Net new quarterly ARR added:ImageYoY growth in quarterly net new ARR added:Image2. $Alphabet(GOOG)$ $Alphabet(GOOGL)$ Google Cloud at a ~$61B run rate growing 34% (not constant currency). Google cloud

For years, every model release followed the same pattern: a flurry of charts showing performance gains across MMLU, HumanEval, GSM8K, and whatever other benchmark happened to matter that quarter. It was a scoreboard for intelligence, a nd every model came with its proof point. A few points higher here, a few tenths lower there. But something changed. Those charts stopped being interesting (at least to me…). Every model sits within a rounding error of each other now, and people seem to have quietly stopped caring. Benchmarks have become saturated.It reminds me a bit of the “index wars” during the early days of search. Back then, Yahoo, AltaVista, and Lycos all bragged about the number of web pages they had indexed, and the bigger the number, the smarter the engine. Then Google came along, a

If we look back over the last few years there are pretty clear patterns of “hot topic debates” that seem to pop back up every so often around AI. One I want to discuss today is the broad topic of “will scaling laws hold.”It’s a nuanced question, because “scaling laws” really mean many things, all of which trace back to data and compute. The debate was broadly could you keep throwing more data and compute at the models to make them better, or would they start to plateau. Over time, nuance has emerged. It’s not just about watching performance scale with more data / compute, but watching performance scale based on where / when / what type of data / compute you throw at models to make them better. Regardless of the nuance, the debate seems to perpetually oscillate between “they wont hold!” to

OpenAI just had their app store moment, and we may look back on it as one of the most significant announcement of the company’s history. The company introduced “apps” inside ChatGPT, and a development platform called Apps SDK to help develops build custom native ChatGPT apps. You can now call an app by name, “ $Spotify Technology S.A.(SPOT)$ , make me a playlist,” “ $Expedia(EXPE)$ , find me a flight,” “ $DoorDash, Inc.(DASH)$ , order my usual,” and ChatGPT will handle the interaction in-line, blending natural language, APIs, and lightweight interfaces into a single conversational flow. It’s a simple but profound shift: ChatGPT is no longer just a reasoning engin