In Part 1, I described two elite engineers of equal ability sitting in the same room, their output diverging not because of talent or effort but because of what they can spend. Token access as infrastructure. Tokens as the new broadband.

Now move down the economic ladder. Far down. Meet Sunita. She runs a small textile trading business out of Nagpur, a Tier 2 city in central India. She has eight employees, a warehouse she shares with two cousins, and a mobile phone she bought eighteen months ago for 11,000 rupees. She uses a free AI chatbot occasionally to draft supplier emails and check tax calculations. It helps at the margins. But it does not do much that she could not do herself with time.

She has heard, vaguely, that bigger companies are running AI agents that handle entire procurement workflows, flagging pricing anomalies in real time, negotiating terms, filing compliance documentation automatically. She has no pathway to get those tools. The subscription tiers that would give her access to agentic AI features are out of reach — they start at prices that represent a meaningful fraction of her monthly margin. So she stays where she is. And the gap between what she can do and what her better-capitalized competitor in Mumbai can do quietly widens, one month at a time.

Why It Actually Matters

It is worth pausing to ask: does frontier AI actually change outcomes for someone in Sunita’s position, or is this just a story about productivity tools for people who were already doing fine?

The answer is that the gap between a basic AI assistant and a capable reasoning agentic model is not a gap in convenience. It is a gap in what problems you can even attempt to solve.

A basic model, the kind available free via a mobile browser, handles discrete tasks adequately: translate this sentence, draft this message, explain this concept, summarize this paragraph. What it cannot do is reason across a business. It cannot ingest three months of purchase orders, identify that one supplier has been overbilling on a specific fabric grade, cross-reference that against market prices, calculate the cumulative loss, and draft a negotiation brief in Marathi, all as a single coherent chain of work. That kind of multi-step, context-holding, judgment-exercising workflow is what frontier models and agentic tooling make possible. It is also precisely the kind of work that consumes expert time in any small business.

For Sunita, access to that capability does not mean a marginal improvement in how she writes emails. It means she would get, for the first time, something close to what a large company has always had: analytical capacity she cannot afford to hire. A capable AI agent could audit her supplier relationships, flag her GST tax exposure, model the cash flow implications of a new contract, and help her prepare for a bank loan conversation in an afternoon. Without it, those tasks either do not get done, or they get done slowly and partially, or she pays a consultant she can barely afford.

Multiply Sunita across the hundreds of millions of small business owners, farmers, health workers, and tradespeople across India who sit in a similar position. The Token Divide is not a tech industry problem. It is a development problem at population scale. It currently applies to some 90% of India’s 600 million smartphone users.

Three Ways to Fix It

There are three plausible paths to closing the Token Divide for the priced-out majority. Each has a fundamentally different logic, and a fundamentally different set of problems.

The first is a right to compute. Governments could treat access to meaningful AI capability the way they treat access to water, education, or electricity: as a public good worth subsidizing. India is already moving in this direction. The IndiaAI Mission, launched in 2024 with a ₹10,372 crore ($1.2 billion) outlay, has deployed over 38,000 GPUs at subsidized rates of roughly ₹65 per hour, with up to 40% subsidies available for MSMEs, startups, and academic institutions. India’s AI Governance Guidelines specifically recommend “targeted incentives such as tax rebates, AI-linked loans through SIDBI and Mudra, and subsidised GPU access for MSMEs.” The intent is clear: break the concentration of AI capabilities before it hardens into a permanent structural divide.

It is an important start. But subsidized AI training compute for startups and researchers is not the same as subsidized access to agentic workflows for Sunita. The math of providing meaningful agentic AI access to hundreds of millions of small business operators, at the usage levels that would actually change their outcomes, is a different order of magnitude entirely. India and other countries where the Token Divide will bite hardest do not have unlimited fiscal capacity even if the digital infrastructure is in place. Right to compute is a compelling principle. As policy, it remains aspirational for the lower-income populations who will need it most.

The second is ad-supported access. OpenAI launched ads in ChatGPT in February 2026, initially for free-tier users in the US, with stated intent to expand globally. The logic is familiar: the internet ran on this model for twenty years. OpenAI’s own framing is explicit that “ads support our commitment to making AI accessible to everyone.” If attention subsidizes access, maybe the Token Divide closes the way the content divide did.

The problem is that the ad model works well for content consumption, where interactions are short, frequent, and easily matched to an ad impression. It fits poorly at best with the kind of deep, multi-step agentic work that actually matters for the Sunitas of the world. An AI agent that runs for hours, ingesting business records and producing a supplier audit, does not slot into any known advertising model. OpenAI is currently targeting $2.5 billion in US ad revenue for 2026, concentrated on short conversational exchanges with high-intent consumers. Funding sustained agentic workflows for MSME operators in Nagpur through that same model is currently hard to imagine.

A third and actually promising fix is edge AI compute. This is the most structurally interesting path, because it does not depend on government budgets or advertiser economics. It depends on hardware that is already in people’s pockets — smartphones.

Here is the insight: the phones people in cities and even rural areas already own are increasingly capable of running AI models locally. A single mid-range smartphone can run a small but useful AI model without touching the cloud. But a single phone’s capability is limited by its memory and processing power. A small group of phones, connected over a local WiFi network, can host and run a substantially more capable model than any one of them could manage alone. The pool vastly exceeds the capability of any single device — though coordination overhead means performance is lower than the raw sum of the parts.

Sunita’s building has eleven people in it. Her brother-in-law runs a transport brokerage two floors up. The accountant across the lane has a newer phone. Five of those devices, loosely cooperating, can do what none of them could do alone.

This is not a distant prospect. Researchers are already working on the core mechanics. Petals (BigScience, 2022) was the foundational proof that volunteers could pool consumer devices to run models too large for any single machine. EdgeShard demonstrated the same principle on mobile and edge devices specifically. Exo, an open-source framework, enables automatic device discovery and larger model distribution across phones on a local network. And SharedLLM, an India-first open-source project currently in early development, is building exactly this stack with non-profit governance and the widest possible hardware compatibility as explicit design goals. These projects are initial indications that the hardware problem is solvable at scale. What does not yet exist is a production-grade coordination layer: software that handles user signup, power and resource management, privacy guarantees, and the hard reality of phones that drop in and out of the network unpredictably. In other words, the hard part is the software — and that’s good news.

The Software That Needs to Be Built

The Token Divide will not close on its own. The economics of frontier AI, high compute costs passed on in token pricing, naturally concentrate capability at the top of the economic pyramid. That is not a moral failure of the AI industry. It is just how costs of data centers work.

The government subsidy path requires political will and fiscal capacity that most affected countries do not have, even if India is demonstrating a credible early attempt. The ad-supported path is live but has no obvious application to agentic use cases. The edge compute path is technically feasible but requires coordination software that nobody has built yet. This is the opportunity for India.

That coordination layer, the software that turns a neighborhood’s idle phones into a shared AI resource, will not be a hardware problem. The hardware is there and getting better every day, in a billion pockets across the markets where the Token Divide bites hardest. It is not an AI model problem. The models that could work on a group of phones are good enough and improving. It is an incentive and orchestration problem, which is exactly the kind of problem that a well-designed software business can solve. The direction of travel is clear: Akamai has argued publicly that centralizing AI in data centers is no longer sufficient, and Nvidia made a $1 billion bet on Nokia to build AI-native mobile networks for distributed edge inference. Moving this from research labs to production and mass distribution is a fantastic opportunity for an innovative software startup, a next-gen router company, or a giant telecom provider, or many companies in between.

The Token Divide is wide, but it can be narrowed. The most interesting response to it may not come from a frontier AI lab or a government ministry. It may come from a WiFi router in a building in Nagpur, quietly stitching together compute on phones to create data mini-centers consuming no water and little incremental power. The idea is less far-fetched than it sounds. If just 10 percent of the smartphones sitting idle across India’s seven largest metros were pooled over local WiFi, they would collectively provide tens to hundreds of exa-operations per second of mobile AI inference capacity. That is not an NVIDIA H100 supercomputer like xAI’s Colossus in Memphis, the world’s largest AI facility. It draws 250 megawatts from the grid to achieve what aggregated smartphones cannot ever match on raw throughput. But they do not have to. Their advantage lies elsewhere: enabling neighborhood- and city-scale AI services using compute that already exists. That is a serious infrastructure proposition for narrowing the Token Divide.

For the first fifty years of computing, the primary limiting resource in software was human capacity. To build something meaningful, you needed an engineering team with the right skills, hardware powerful enough to run your system, and the organizational discipline to specify requirements, write code, run tests, deploy to production, and keep the whole thing humming along. That chain was long, expensive, and slow. The “software crisis” of the 1960s and 1970s was essentially a crisis of that chain: projects ran over budget, fell behind schedule, and failed at alarming rates, not because the ideas were bad, but because the execution machinery was so hard to assemble and operate.

Each decade brought tools that compressed one link in the chain. Structured programming in the 1970s. Object-oriented languages in the 1980s. The open-source movement and agile methods in the 1990s and 2000s. Smartphones in the 2000’s. Cloud infrastructure in the 2010s. Each of these offered a genuine leap in productivity, deployment, and scaled access to software-based solutions. None of them changed the fundamental constraint: you still needed skilled engineers, and a lot of them, to build and maintain anything serious.

That constraint has now been broken. Not reduced. Broken.

Today, a single developer with the right tools and sufficient budget can spec, build, test, and deploy a working system in days or weeks that would have required a team of engineers months or years to produce even five years ago. The bottleneck has shifted. It is no longer about having the engineering resources to build and deploy. It is about having the cognitive clarity to define the right problem, the judgment to verify that the AI-generated solution actually solves it, and the token budget to run the process at scale. Two of those three are fixed costs, approaching free. The third is the opposite — expensive and variable.

This shift has created a new divide. I call it the Token Divide. And unlike the digital divide of the 2000s, which separated people across the globe by access to connectivity, this one is 100 to 1,000 times more consequential. It is happening now, not in some hypothetical AI future.

This article is Part One. It starts where the gap is most visible and most dramatic: elite software and design engineers. Part Two will look at the same problem at the bottom of the income ladder.

What Elite Looks Like Now

In February 2025, former OpenAI researcher Andrej Karpathy coined the term “vibe coding” to describe a new mode of software development: engineers building things by simply chatting with AI models. The tech press and even many experienced software developers treated it as a clever novelty. It was actually a signal flare.

Since then, a subset of developers has gone far beyond vibe coding. They are using tools like Claude Code, Cursor, and agentic loops that chain AI models together, with one agent directing others, one agent reviewing another’s quality and security, running tests, catching regressions, and iterating, all without a human in the loop, for extended stretches, from overnight to multiple days. The output is not “pretty good AI-assisted code.” Cutting-edge developers are creating production-quality software, produced at a velocity that no human team could match.

One tech CEO described his experience as being able to “finish a year’s worth of work in a week” using Claude Code. He estimated a five-times productivity increase. This is not an isolated anecdote. An Amazon engineer recently described the 3x productivity gain he is seeing and put it in clear recent historical context on VentureBeat. Y Combinator reports that 25% of its recent startup cohort has codebases that are 95% AI-generated. These are not hobbyist projects. They are funded companies shipping products to paying customers.

The Stack Overflow 2025 Developer Survey, drawing on over 65,000 developers globally, found that 84% of professional developers now use or plan to use AI coding tools, up from 76% the year before. Adoption is not the question. The question is how expansively people are using AI, and whether they have the token budget to go deep.

The Previously-Hidden Variable: Token Budget

Here is the thing that most coverage of the AI productivity story misses. The engineers achieving multiplicative gains are not just more skilled or more motivated. They are spending real money, in many cases thousands of dollars per month on AI tokens, to do it.

Serious agentic coding workflows burn tokens fast. Uber’s CTO recently noted that the company spent its entire annual AI budget in 4 months. According to developer productivity benchmarks compiled in 2026, teams using agentic AI tools spend $200 to $2,000 or more per engineer per month in token costs, on top of per-seat licenses. Looping multiple models for review and testing can compound that further. For an individual developer going deep, the monthly spend is now a meaningful percentage of their salary.

Jensen Huang, CEO of Nvidia, put it bluntly (and obviously self-servingly), saying Nvidia engineers should be consuming AI tokens worth roughly half their annual salary each year to be fully productive. Half their salary. Let that sink in. For a senior engineer earning $200,000, that is $100,000 in annual token spend. Most individual developers and most companies are nowhere near that level, not because they lack the desire, but because they lack the budget.

Anthropic’s recent pricing moves underscore the structural reality. Frontier models like Claude Opus are not cheap. Agentic use cases that loop models together multiply that cost. The economic floor for operating at the leading edge is real and rising.

Two Engineers, One Winner

Imagine two engineers. Same training. Same intelligence. Same drive. Same GitHub history. One has a $2,000-per-month token budget and six months of experience working with agentic coding tools. The other has a $20-per-month subscription cap and is just getting started.

The first engineer runs Claude Code autonomously against a complex, AI-developed codebase, sets it to generate tests, review its own output, and iterate. She goes to bed. In the morning, it’s ready for a final code review, for which she uses another tool to do as thorough a review as she used to. The second engineer is manually prompting, getting good but not great suggestions, writing 25+% of code, and hitting rate limits mid-task as his paltry monthly budget gets exhausted.

This is not a hypothetical. It is described across developer communities as a “fork,” not a spectrum. The gap between those who have an “automated factory” and those who have a “calculator with a daily usage cap” is not a gap that closes with motivation or study. It is a gap that closes with budget, and with time spent learning how to wield that budget effectively.

The Asia Times captured this well: “The gap being built is not about who has the better engineer. It is about who handed their engineer an automated factory vs. who handed theirs a calculator with a daily usage cap.” Every autonomous workflow completed, every codebase refactored overnight, widens a moat that talent alone cannot close.

Meanwhile, a piece in Medium by Aseem Rajvanshi makes the compounding dynamic explicit: companies with deeper AI budgets “move so much faster that they can afford to waste tokens on experimentation, rapid prototyping, and ambitious architectural decisions that would bankrupt their budget-constrained competitors.” The best engineers are increasingly choosing to work only for employers who will not put a ceiling on their AI tool usage. Why would they not?

Tokens as Infrastructure

We are watching token access become what broadband was in 2003, except faster and with steeper slopes. A TechCrunch analysis observed that AI tokens are emerging as a “fourth pillar of engineering compensation,” alongside salary, equity, and benefits. Some leading tech companies are already offering token budgets as a recruiting tool. “Developers who don’t use AI assistance are increasingly seen as inefficient,” one analysis noted. “Token access isn’t a perk. It’s a prerequisite for productivity.”

Google CEO Sundar Pichai has said publicly that “the digital divide cannot be allowed to become an AI divide,” calling for investment in computing infrastructure and connectivity to prevent it. He is right to worry. But the divide he is naming is not only about access to AI products as consumers. It is about the capacity to work with AI at full power, as a professional tool, which requires a budget that millions of talented engineers around the world simply do not have.

This Is Not Coming. It Is Here.

The productivity data is being published, and can be confusing. A 2025 randomized controlled trial by METR of 16 experienced open-source developers across 246 real-world tasks found that casual AI users, those using tools intermittently without deep agentic workflows, actually completed tasks 19% slower than the control group. The tools confused them more than they helped.

But that is casual use of AI dabblers. At the other end, Faros Research found that engineers deeply embedded in agentic workflows completed 21% more tasks and merged 98% more code. The spread between the casual user and the power user is not narrow. It is chasm-wide and growing. And those AI users with adequate token budgets will be more productive than any number of those without, thus leveling the field that used to be dominated by companies with exceptional staffing budgets.

We are still in the early innings. Claude Code, Cursor, Codex, and their successors are improving fast. The cost of compute is high, and the economic floor is real. As Anthropic’s Dario Amodei has noted, leading AI companies may need to generate hundreds of billions in annual revenue to sustain the compute costs that frontier models require. That cost gets passed on in token pricing. The gap between what is possible and what is affordable for most developers will not close on its own with current models.

For the first time in software history, two engineers of equal ability, sitting in the same room, doing the same job, can have vast differences in output. Not because of what they know or how hard and smart they work, but because of what they can spend.

What Comes Next

Elite engineers are the most visible edge of this divide, but they are only the beginning of the story. Their problems may be addressed by a combination of falling token prices, model optimizations and company budgets. Soon, edge AI compute, where increasingly capable models run locally on the computers the developers already own, is likely to emerge as a powerful solution. Imagine immense compute power already in existence being aggregated and coordinated by a new layer of software, to provide much lower and increasingly powerful agentic reasoning models that run for a fraction of the cost of today’s models, or in many cases, for “free”.

In Part Two, I will look at the other end of the spectrum, starting with small business owners in Tier 2 Indian cities navigating the need for agentic business solutions they cannot afford. For them, the Token Divide may ultimately matter more than it does for any developer in a startup enclave, as the stakes there are not productivity multiples. They are local and regional competitiveness, livelihood, food security, and economic inclusion. Opposite ends of the economic pyramid will experience the same token divide. Unless good solutions emerge that can operate at scale, I will propose one.

Executive Summary

SpaceX went public on June 12, 2026, raising $75 billion at a $1.75 trillion valuation — the largest IPO in history. The stock closed ~20% higher at the end of trading that day. A significant portion of that valuation — maybe half — rests on the orbital AI data center thesis: satellites running AI compute workloads in space. I’m not an expert on space or data centers or physics, but I know how to ask hard questions, and I have a few smart friends to ask, and the answers just don’t add up.

Retail investors bid the stock up due to market hype and small supply — that’s not surprising given Musk’s track record. The bigger question is what drove smart institutional investors to pay the IPO price in the first place? Why would they ignore Morningstar’s assessment of the value being 55% LESS than the target IPO price?

This document examines the physics, the economics, the market mechanics, and the public market investment drivers. I did this research with the goal of learning more about the nuanced dynamics of public markets and IPOs, as I’m supporting two of our portfolio companies in India, Betterplace and DriveU, that are on an IPO journey for 2027. Note that contrary to nearly all of my posts, the vast majority of this post is AI-written, based on lots of prompting back and forth by me. I used Claude and then had Gemini fact-check it. So there’s a good chance even the math is correct 🙂

The overall thinking and hard questions are definitely mine, so I own that and hope you find the thought process helpful. I also hope this piece will help my readers understand two things: 1/ one can’t take IPO CEO’s comments and prognostications for the future at face value, however much you might presume the market would reject outright impossibilities, and 2/ the markets sometimes have reasons that are not obvious to buy a stock promoted by a founder making dubious claims if he/she has accomplished very big things in the past. As Elon has.

Core findings and realizations: The physics of heat rejection in space create hard constraints that SpaceX's published AI1 specifications do not remotely satisfy. The radiator math simply does not work. This is arithmetic, not opinion — but the market ignored the facts in facor of the overall narrative and option value in other areas, and may not price the lack of viability of AI1 for years.

Net: narratives and option value apply in public markets as well as VC markets, even when part of the $800B of narrative option value is physically impossible or uneconomic or both.

1. The Thermal Physics Problem

1.1 Why Cooling is a Central Constraint

A ground data center sheds waste heat into air and/or water. A satellite has neither. It can only radiate heat as infrared light — governed by the Stefan-Boltzmann law, a fixed physical relationship that cannot be engineered around.

The equation: Q = ε · σ · A · T⁴

where σ = 5.67×10⁻⁸ W/m²K⁴, ε ≈ 0.85–0.91 for good space radiator coatings, and T is the radiator surface temperature in Kelvin.

1.2 Radiator Area Required to Cool 150 kW (Musk’s stated power of AI1)

The following table shows how much radiating surface area is required to reject 150 kW of waste heat at various operating temperatures:

The International Space Station is the gold standard for large-scale heat dissipation. At the ISS-proven operating range of ~300K, roughly 370–430 m² of radiator surface is required. The ISS itself rejects about 70 kW across 422 m² of radiator — at ~166 W/m². That’s half the stated power of AI1.

1.3 What SpaceX Claims for AI1

SpaceX’s AI1 satellite specifications, revealed June 8–9, 2026:

• Peak compute: 150 kW; sustained: 120 kW

• Radiator area: 110 m² deployable liquid radiator

• Claimed radiator performance: ~1,400 W/m²

• Total satellite mass target: ~2.14 tonnes (at 70 kW/tonne efficiency claim)

• Wingspan: 70 meters (wider than a Boeing 747-8)

• Orbit: 600 km LEO

1.4 Why the Numbers Don’t Work

Critical problem: The ISS, the largest thermal control system humanity has ever flown, achieves ~166 W/m². SpaceX claims 1,400 W/m² — eight times that performance — as a baseline assumption.

To achieve 1,400 W/m², the radiator surface must operate at approximately 470K (~200°C). But chips must run way cooler than that. The liquid cooling loop must move heat from ~85°C chip junction temperature up to a 200°C radiating surface — which requires a heat pump, which consumes power, which generates more waste heat, compounding the problem.

Working backwards from physics:

• A 110 m² radiator at chip-safe temperatures can reject roughly 110 × 166 W/m² = ~18 kW at ISS-proven performance

• To reject 150 kW at ISS-proven density, you need ~900 m² of radiator

• Even optimistic engineering improvements suggest ~138 m² minimum before real-world losses for manifolds, plumbing, and micrometeoroid shielding

• Maybe there is unspoken potential help from putting radiators on the back of solar panels like what the James Webb telescope does. This is not simple engineering, but it is known. If so, the math might not be quite so bad on the dissipation front, but there are still the other minor issues of weight, communication latency, degradation of GPUs etc.

1.5 The Mass Budget Problem

SpaceX claims 70 kW per tonne for AI1. At 150 kW peak, the entire satellite — solar arrays, radiator, structure, propulsion, comms, shielding, and GPUs — must weigh no more than 2.14 tonnes.

A carefully designed aluminum space radiator runs approximately 5 kg/m². At 110 m², the radiator alone weighs ~550 kg — already 26% of the entire mass budget before a single GPU, solar panel, or structural member is added.

This is not an engineering challenge that better materials or more funding resolves. Stefan-Boltzmann is a law, not a guideline. The surface radiates what the physics allows at a given temperature, no more.

1.6 Elon’s Comparison to Starlink

SpaceX has not publicly released exact specifications for Starlink v2 mini satellites, but photometric and other analyses estimate a solar panel area of approximately 105 m². Exact specifications have not been publicly released by SpaceX.

What can be inferred:

The v2 mini satellite has a launch mass of approximately 740 kg. Based on the ~105 m² solar panel area and a typical solar panel efficiency of around 30% (likely overstated) at 1,360 W/m² incident, the v2 mini harvests roughly 20–28 kW of electrical power. Since virtually all electrical power becomes waste heat, the satellite must reject somewhere between 10 and 28 kW depending on actual panel efficiency and system needs, likely via a 4x4 meter passive radiator.

At an ISS-proven radiator density of 166 W/m², that implies roughly 120–170 m² of radiator area — which seems implausible for a 740 kg satellite given mass constraints. In practice, Starlink satellites run the radiators hot to reduce area. They can do this because communications payloads are far more thermally forgiving than GPU compute. The actual radiator is almost certainly a modest body-mounted panel, not a deployable system.

The critical comparison to AI1:

This is where it gets revealing. One analyst used ISS radiator mass density of 15 kg/m² as an upper-bound estimate, noting that if satellites require additional radiators for a full 28 kW thermal load, this could add approximately 660 kg per satellite, bringing total mass to roughly 1,775 kg — and that should be treated as an upper bound since ISS technology is dated.

So a Starlink v2 mini rejecting ~28 kW already pushes mass budgets hard. AI1 needs to reject 150 kW — more than five times as much — from a satellite in the same mass class. That ratio is the clearest single articulation of why the AI1 thermal spec doesn’t close: SpaceX’s own most advanced production satellite, optimized over years, handles roughly 20–28 kW of heat rejection. AI1 claims 5x that from a comparable mass platform, using radiator technology that has never been demonstrated at scale.

2. The Terrestrial Data Center Comparison

2.1 What 150 kW Costs on the Ground

At 150 kW, a facility sits in edge/small enterprise data center territory. The following construction costs exclude GPUs and servers:

A terrestrial 150 kW AI data center shell — power, cooling, building, connectivity, no GPUs — costs roughly $2–3M in the US today.

2.2 The Infrastructure Cost Gap

For comparison, the ISS external thermal control system rejects ~70 kW across 422 m² at an estimated cost of up to $500M for that system alone. Even at aggressive manufacturing scale and Starship launch economics, each AI1 unit almost certainly costs $50–100M+ before a single GPU is installed. That represents a 20–50x cost premium over a terrestrial build for the infrastructure alone — before accounting for the 5-year deorbit cycle, no serviceability, radiation-induced hardware degradation (roughly 20–30% performance loss per year in LEO), and latency constraints. Depending on way too many variables and counting on SpaceX performing their demonstrated prowess in reducing costs of satellites (and their lift engines), that $500M could be substantially lower, but it’s hard to imagine it being remotely on par with terrestrial costs.

2.3 Topics Not Addressed on the Ground vs. in Space and on the Moon

The following are not small things, but I didn’t have time to get into all of them. There’s a great [long] video discussing all the issues here:

• Communication latency and overall speed - massive disadvantage in LEO, likely insurmountable disadvantage on the moon (2+ second round trip latency).

• Degradation of silicon due to cosmic radiation and cost of shielding and/or hardening of the silicon to radiation. Some reports suggest 80% of capacity will be gone within the 5-year projected lifespan while in LEO. On the moon, way worse.

• Orbital congestion.

• Moon: 15 days of sunlight, 15 days of dark.

3. The Institutional Bull Case

3.1 What the Market Actually Believes

The orbital data center thesis is not the only reason to own SpaceX. Sophisticated institutional investors buying the IPO likely underwrote a layered thesis:

Layer 1: Starlink Is the Anchor

Starlink generates 61% of 2025 revenue and is the only profitable SpaceX segment. A standalone Starlink business — proven, growing, near-monopoly in LEO broadband — is worth $50–100B on conventional comps. This is the institutional foundation.

Layer 2: The Grid Bottleneck Is Real

Terrestrial data center deployment is genuinely bottlenecked by 5–7 year power grid interconnection queues. The $2–3M build cost comparison is misleading if you cannot get power for years. Space sidesteps permitting, grid queues, and community opposition — a timing arbitrage that has real economic value even if it is not permanent.

Layer 3: The Launch Cost Curve

At SpaceX’s target of $200/kg with Starship (versus ~$2,700/kg on Falcon 9), the economics of orbital infrastructure change fundamentally. McKinsey analysis suggests $500/kg is the threshold where space compute becomes cost-competitive with terrestrial alternatives. IFF Starship hits $200/kg, the physics constraints become engineering problems worth solving at scale.

Layer 4: Revenue Already Exists

Anthropic is reported to be paying SpaceX approximately $1.25 billion per month to rent xAI data center space. Google has agreed to pay roughly $920 million per month for AI capacity. This is not speculative demand — there is an existing revenue base that orbital compute extends.

Layer 5: Option Value Pricing

Institutions appear to be pricing the delta between Starlink/launch value (~$600–800B) and the $1.75T IPO price as the cost of a portfolio of options: orbital compute, Starship point-to-point travel, Mars infrastructure, Terafab chips. You don’t need all of them. You need one to be real enough over a 10-year horizon.

4. S&P Index Inclusion: An Error and Its Correction

ERROR in original Claude.ai analysis (note from Will - always read AI output with your critical thinking brain in full gear!): This analysis initially stated that institutional investors were buying SpaceX partly as a front-run on S&P 500 index inclusion forced buying — implying a large mechanical catalyst. This was materially wrong.

4.1 What Actually Happened

S&P Dow Jones Indices opened a consultation in May 2026 proposing to reduce the seasoning window and waive the GAAP profitability requirement for megacap IPOs. On June 4, 2026 — one week before the IPO — S&P rejected its own proposal entirely.

CORRECTION: All existing S&P 500 eligibility criteria remain unchanged: 12-month seasoning requirement, four quarters of positive GAAP earnings, and minimum investable weight factor. S&P stated explicitly that exceptions should not be granted solely based on market capitalization. SpaceX cannot enter the S&P 500 until at least mid-2027, and only if it achieves GAAP profitability.

4.2 The Profitability Problem

SpaceX moved from a profit of $791M in 2024 to a loss of $4.9B in 2025, with an additional $4.3B loss in Q1 2026 alone. The xAI acquisition and Starship development are driving deep losses. At this rate, GAAP profitability — the S&P 500 gating requirement — is not imminent.

4.3 What Is Real: Nasdaq-100 Inclusion

Nasdaq changed its own index rules effective May 1, 2026, creating a ‘fast entry’ path for companies ranking in the top 40 by market cap. SpaceX becomes eligible for Nasdaq-100 inclusion approximately 15 trading days post-IPO — around July 7, 2026. This creates real forced buying from QQQ and Nasdaq-100 tracking funds.

However, the Nasdaq-100 tracks approximately $6T in assets versus the S&P 500’s ~$45T. The mechanical forced-buy catalyst from Nasdaq-100 inclusion is real but materially smaller than the S&P 500 event that was incorrectly assumed in the original analysis.

The corrected picture: Nasdaq-100 inclusion creates a known July 7 forced-buy event, but the large S&P 500 passive buying thesis that institutional investors may have been sold simply does not exist under current rules. The CNBC IPO consultant comment is apt: 'There is zero math that makes any sense whatsoever.'

5. What a Physics-Constrained Investor Would Actually Buy

5.1 What Space Genuinely Offers

Physics gives three real advantages to orbital infrastructure:

• Solar power: ~1,360 W/m² incident, ~250 W/m² harvested (max). Real, proven, scales linearly with array area.

• Radiative cooling: real but constrained. 166 W/m² is the proven ISS baseline, not 1,400. Maybe it comes out in the 500+W/m² range if some unforeseen improvements in engineering get around the physics.

• No permitting or power grid queue: a timing arbitrage for sure, not a permanent structural advantage given the cost deltas.

5.2 The Workload That Actually Makes Sense (Hint: Not Datacenters)

The physics-compatible orbital compute use case is Earth observation data processing at the edge (aka in space), not general AI training or inference.

Why it survives scrutiny:

• Latency is irrelevant: you are processing sensor data collected in situ, not serving user requests

• Compute requirements are modest: image processing, change detection, SAR processing runs on far less than 150 kW, making radiator size tractable at proven ISS-class densities

• The alternative is genuinely inefficient: raw data downlink, ground processing, result uplink is bandwidth-constrained

• Defense and intelligence customers pay extraordinary premiums for reduced latency intelligence processing

• A 20 kW edge processing satellite at proven radiator densities is physically achievable today. This is not AI1. It’s an entirely different value proposition.

5.3 The Real Starship Investment Thesis

The rational Starship-enabled thesis is not a new physics-defying compute paradigm. It is simply: things that already demonstrably work in space get dramatically cheaper, and some marginal use cases that were previously uneconomic become viable. Remote sensing, communications relay, Earth observation — all benefit from launch cost reduction without requiring unproven thermal performance.

6. Trading Implications

6.1 Why a Short would be Dangerous Despite Physics Being Correct

The physics critique is likely correct. Nonetheless, a short trade is dangerous:

• Float is 4%: an extraordinarily dangerous borrow situation. A squeeze requires only a narrative catalyst, not a fundamental one.

• Nasdaq-100 inclusion July 7 creates a known forced-buy event with no price sensitivity.

• Loss on a short is theoretically unlimited: a further 50% run to $2.5T is entirely plausible on narrative momentum.

• Musk controls the news cycle: one Starship launch or government contract announcement reframes the narrative for weeks.

6.2 Why the Market Won’t Price Physics Quickly

Knowing the physics is wrong and having a tradeable thesis are completely different things. The people with domain knowledge to immediately see the radiator math doesn’t close — aerospace engineers, semiconductor thermal architects — are almost entirely not the people with capital to express the trade. Sell-side research from Goldman and Morgan Stanley is conflicted on a deal they underwrote. Price discovery will come from:

• AI1 prototype telemetry in 2027: actual thermal performance data from orbit is inarguable

• Quarterly earnings: compounding losses against the S&P 500 inclusion gate forces institutional reassessment

• A credible short report from a firm with sufficient distribution to shift narrative

The market is not particularly efficient at incorporating engineering reality into narrative stocks on short timeframes. It is quite efficient at incorporating it once the hardware either works or doesn’t in public.

7. Conclusion

The SpaceX IPO at $1.75T is simultaneously one of the most defensible and most questionable valuations in market history. Starlink is a real, profitable, growing business worth tens of billions on its own. The launch monopoly is genuine (today). The Starship cost curve, if it materializes (which has not been proven at all yet), changes orbital economics fundamentally.

But the orbital AI data center thesis — the portion of valuation that Wall Street is being asked to underwrite most aggressively — requires radiator performance eight times better than anything humanity has demonstrated in space, in a satellite weighing 2.14 tonnes total. The S&P 500 forced-buying thesis that may have been used to sell institutional participation does not exist under current index rules.

What Wall Street bought was a narrative dressed as a structural trade. The physics critique is correct. It may take years and a failed prototype to matter to price.

The rational investor buys the Starlink thesis and treats orbital compute as a free option on Musk being right about physics he hasn’t yet demonstrated. The disciplined skeptic waits for options to list, lets volatility compress, and buys time — not proximity to the current price.

Appendix: Understanding the Valuation Implied by Each Scenario

SpaceX IPO market cap: $1.75 trillion at $135 per share.

• $775B (-56%): Starlink at full standalone value (~$80-100B) plus Starship launch monopoly. Orbital compute valued at zero. Full reflection of physics constraints.

• $1.1-1.2T (-33-37%): Starlink plus launch monopoly plus partial Starship optionality. Orbital compute premium stripped out. Reachable on a launch slip disclosure alone.

• $1.25-1.35T (-26-30%): Narrative dented but not broken. Requires IV spike to achieve 2x on the put. Needs a sharp, fast move rather than a slow drift.

For context, the S&P 500 inclusion thesis — which requires four quarters of GAAP profitability after recording $4.3B in losses in Q1 2026 alone — is the single largest institutional justification for the current valuation premium above $1.1-1.2T. A market consensus that profitability is 3+ years away would itself justify the $1.1-1.2T repricing scenario without any hardware failure.

In my daily scan of AI news, I keep reading about the antipathy Gen Z feels toward AI. The most recent example was Eric Schmidt, former Google CEO, receiving a cold reception at his 2026 commencement address at the University of Arizona. Empathy was never Eric’s strong suit. I noted that when I worked for him nearly 40 years ago; his brilliant success at Google hasn’t led to change…

If you’re Gen Z, your thinking goes something like: AI is taking our jobs, it’s horrible for the environment, it makes us dumber, and probably worse. The surveys are unambiguous about the fears, but the research on actual job impact is genuinely contested. And there is simply no way to put the genie back in the bottle.

So what’s a Gen Z to do? My answer is easy to say but less easy to execute: accept some of the inevitable, reject what shouldn’t be tolerated, and take agency into your own hands. Let me lay out some foundational context, then give you seven pragmatic ways to cope with and ultimately flourish alongside AI.

The Mood Has Shifted Fast — You Are Not Alone

The April 2026 Gallup/Walton Family Foundation “Voices of Gen Z” survey of 1,572 Americans aged 14–29 found that 31% now feel anger toward AI (up nine points from 2025), while excitement dropped 14 points to 22%. Among Gen Z already in the workforce, 48% say the risks of using AI at work outweigh the benefits. And here’s a particularly painful wrinkle: 47% of Gen Z workers hide their AI use from employers out of fear that demonstrating AI proficiency signals their own replaceability. The level of AI unpopularity is growing and surprising to tech leaders. If AI were a political candidate fielded by the tech industry, it would be losing every election.

Can We Slow This Down?

No. A major turning point came in 2023 when Meta and Mistral released highly capable open-source AI models that are free for anyone to download and run. There’s no practical way to regulate open-source technology at this level of capability. The recent 2026 drops from major Chinese labs are nearly as capable as the most advanced models from OpenAI, Anthropic, and Google. And even if we could regulate the tech, the current body politic does not appear to be seriously trying. The genies are out of the bottles and reproducing. There is no way back.

Job Market Softness Is Real

The Class of 2026 is entering the worst non-pandemic graduate job market in over a decade. Recent-grad unemployment (ages 22–27) in the US sits at 5.7% versus 3.25% pre-pandemic. Employer hiring projections are up just 1.6% compared to 16.6% in 2019. The deterioration is concentrated in white-collar entry-level positions.

Is AI to blame? The evidence is genuinely contested. A Stanford study found a 13% decline in employment in AI-exposed occupations among workers aged 22–25 since 2022, with a nearly 20% drop for entry-level software developers specifically. Goldman Sachs estimated AI substitution is displacing roughly 16,000 jobs per month, net. But skeptics at the Economic Policy Institute note that unemployment is rising for young workers without degrees too, suggesting macro forces — post-pandemic tech rationalization, geopolitical uncertainty, corporate caution — deserve at least equal blame.

The best answer i can come up with after hours of reading studies is that AI is a meaningful contributor, but is almost certainly not the only driver. On top of that, the situation varies market to market and by industry. In all cases, it’s likely to get worse before it gets better. And AI's impact on entry-level jobs is absolutely not going away.

Seven Things You Can Actually Do

1. Move from Denial and Anger to Acceptance (But Not Forgiveness)

You’re allowed to be angry. But anxiety and hostility won’t slow AI down any more than denying climate change has slowed carbon emissions. In their book Genesis, Schmidt, Kissinger, and Mundie compare AI’s arrival to discovering alien life, capturing both its non-human strangeness and its civilization-altering significance. The alien is here. We need to learn to live with it, on our own terms.

2. Do Not Accept the Hegemony of the AI Industry

Accepting the inevitable doesn’t mean accepting everything. A few concrete moves: A/ Vote for leaders who will push back on AI’s environmental costs; many are already protesting the data centers guzzling water and driving up energy prices. These are appropriate targets for near-term regulation and aggressive taxation. B/ Support local data regulations that mandate authentic labeling of AI-generated content and combat deepfakes. C/ Think hard about AI-driven wealth concentration: if AGI restructures the economy, make sure you’re backing people who will fight for a fair distribution of the gains. D/ Do you best to work for companies that take a humanistic approach, using AI to augment humans, not replace them. E/ Pick your AI tools based on values alignment. The models are increasingly commoditized; you can vote with your wallet.

3. Learn, and Learn to Love Learning Even More

Everyone in the AI industry disagrees about many things, but they all agree on one: the pace of change is unprecedented and nearly impossible to forecast even a few years out. “Lifelong learner” has taken on a new and urgent meaning. You also have to be ready to unlearn things that were true six months ago. One reason I started in VC was for the continuous learning opportunities; VC + AI is that on steroids…

4. Don’t Wait for the Perfect Career Path to Emerge

With AI changing what’s valuable in the job market almost continuously, trying to plot a perfect career arc right now is mostly an exercise in frustration. Any first job with any reasonable manager is a fine starting point, regardless of how well it fits your vision of your future. In his excellent book Range, David Epstein shows through well-documented stories that many exceptionally successful people took surprisingly circuitous routes to get there. I used to say in my bio: “If you don’t know where you’re going, any road will get you there”. Meaning you should know, or you’ll wander. I’ve updated my view. It’s okay not to know. Just keep learning, work hard, and stay in motion. Here are some specific ideas on how to go after that first job.

5. Focus on Skills That Are Uniquely Human (For Now)

In our upcoming book, Flourishing with AI, my co-author Ravi and I devote an entire chapter to skills that hold up in an AI-transformed world. The short version: develop adaptability, critical thinking, empathy, and teamwork skills in school. Then build more judgment, resourcefulness, and storytelling abilities as your career progresses. These aren’t easy skills. In an AI era, they’re the hardest ones to automate and therefore the most valuable. You can discuss these skills with our AI chatbot that has a draft of the book, here.

6. Learn What Talent Recruiters Actually Look For

I recently spoke with a senior partner at Egon Zehnder, one of the top global executive recruiting firms. They focus on four attributes at every level of their hiring: Determination, Insight, Engagement, and Curiosity. Determination isn’t just ambition; it’s the ability to recover from setbacks, course-correct, and sustain energy through uncertainty. Insight, in an age of information overload, means synthesizing competing signals and converting ambiguity into judgment. Engagement means building trust and rallying people through change. Curiosity means staying hungry for new ideas. Every one of these applies whether you’re 22 or the CEO of a Fortune 100 company. Start cultivating them now.

7. Use AI Aggressively, But Guard Your Brain

A widely-cited MIT study found that frequent ChatGPT users showed significantly reduced brain connectivity compared to unaided writers, and critically, struggled to regain cognitive independence when the tool was removed. Cognitive atrophy is not a passive decline; it’s an active loss of capacity. Here’s how to avoid it: generate your own answer before asking AI for one. Regularly solve problems the hard way. Treat AI outputs as evidence to evaluate, not conclusions to accept. Verify claims, test reasoning chains, and develop a clear mental model of where the tool fails. Most importantly, be honest with yourself. When you can no longer answer basic questions in your field without AI, that’s not efficiency, that’s surrender. AI should extend your thinking by handling what you’ve consciously delegated. The moment it starts replacing thinking you didn’t notice outsourcing, you’ve lost something important.

Conclusion

Booing Eric Schmidt and others off the stage may feel cathartic, and it’s clearly understandable for those struggling with job prospects. But it won’t slow AI down. The harder, less immediately satisfying work is what I’ve laid out above: accept what cannot be reversed, refuse what should not be tolerated, and seize agency wherever you still have it. At the ballot box. With your wallet. In the companies you choose. In the tools you use. In what you learn every single day. And most importantly, in the cognitive ground between your ears.

Gen Z didn’t create this mess and shouldn’t have to fix it alone. But every generation has had to navigate technological and economic shocks that arrived faster than the institutions meant to govern them. Yours is no different. Schmidt’s “rocket ship” was the wrong metaphor. This isn’t a ship you board while someone else picks the destination. It’s a machine that will be steered by whoever shows up with both hands on the controls. That should be you. The alien is already here. Get to know it, set the rules of engagement on your own terms, and flourish.

My friend Suraj Moraje, founder of the Eka Fellowship, invited me to lead a workshop on AI with 32 Class 10 kids on the closing day of their program at the Fireflies Intercultural Center, South Bangalore, India. I had a simple goal: convince the kids that AI could be their friend and ally, while informing them of the risks and issues they must be aware of as they use it. I was frankly amazed at how much these kids had thought about AI and impressed by their basic AI fluency. I could easily imagine running the same workshop with middle- and high-income kids at a school anywhere in the world and achieving a very similar level of conversational depth about both what AI does and what kids should do to prepare to enter the workforce in an AI-infused world.

Eka Fellowship: Bridging the Economic Mobility Gap

Before I get into the workshop, a little background on the Eka Fellowship and these kids’ socioeconomic conditions will provide important context. Eka’s mission is to “accelerate seven generations of mobility into one” by providing underserved students with an eight-year, holistic support system that guides them from Grade 9 through to their first job. The kids attend government or low-cost private schools, where they do not get a great education. Their families have a median household income of ₹20,000 / USD ~200 per month. Living on this for a family in Bangalore in 2026 is an exercise in survival. While technically above the official government “poverty line,” the family would be living in what economists call “relative poverty”. Across the 8 years that Eka supports their Fellows, they build five things simultaneously: career clarity, academic capability, professional English, financial resilience, and the emotional toolkit to persist in high-pressure environments.

Likes and Dislikes of AI

We started the conversation by asking the kids what their experiences were with AI. Knowing they had been through an earlier workshop led by Eka, I wanted to get right to understanding their likes, dislikes, hopes, and fears. In italics are verbatim quotes, spoken in clear and confident English (congrats to Eka for your work in this area!)

Things they liked:

1. Emotional support when they’re feeling low: “It’s helpful… like whenever we are feeling low, if we talk to the AI, it will respond in a good way.”

2. Explaining difficult concepts more simply: “If we don’t understand some concept, we can say to explain in an easy way, so that we can understand it and then read it in the book.”

3. Fast research and knowledge building: “It’s a search engine… if I have anything, doubts about science or any general knowledge I need to know about, I will search in the AI… to gain that knowledge, and it’s very good at that.”

Then we got into things it didn’t do well, or maybe couldn’t ever do. I went deeper into this area, as my goal was to get the kids to think about both their own learning in more AI-proof areas and also to be alert to AI leading them astray. The kids were excited to think about and discuss these topics, more than in the previous section:

1. Surgery / medical operations: “The AI can’t come and do the operations on the patients… AI can’t do operations.”

2. Genuine counseling & deep emotional understanding: “I think it’s missing a human quality… when a person talks to another person, they understand… but I think AI is only trying to give a positive answer when people ask when they are feeling low, whereas humans… we understand each other.”

3. Distinguishing right from wrong: “Differentiate between good and bad. It’s the biggest weakness of AI… sometimes a false answer, and it will say it’s correct and it’s true.”

Making AI Their Ally

Taken together, the kids’ comments from this session form a surprisingly mature mental model of AI that I think is instructive for all of us. Most of them see AI as a genuinely powerful helper: it explains difficult concepts in simpler terms, walks them through homework as worked examples, answers questions faster than any traditional reference, and even offers suggestions or encouragement when they’re feeling low.

But they also draw a clear boundary around what remains distinctly human: empathy and real counseling, moral judgment, trust and persuasion, and hands‑on work in the physical world (from teaching to medicine and craft). By the end of the workshop, I think they understood clearly that the right response is neither to reject AI out of fear nor to surrender to it uncritically, but to use it as a learning and productivity amplifier while rigorously double‑checking its outputs, thinking for themselves, and intentionally investing in the human capabilities they believe AI lacks: empathy, creativity, ethical reasoning, real‑world problem‑solving, and collaboration with other people.

For leaders thinking about the implications of AI and “future of work” for all of humanity, this is useful grounding: these young teenagers from low-income families in India are correctly asking whether AI will replace humans, but at the same time, they are also thinking about how to position themselves as humans who plan to work well with AI.

My co-author Ravi Venkatesan and I have lots more to say on this topic, addressing early-to-mid-career professionals as well as senior leaders, in our upcoming book, Flourishing with AI, to be published by Penguin later this year.

An early draft is already available via our chatbot — try it out and tell us what you think!

In February 2026, Anthropic launched Claude Cowork and Wall Street lost its mind. We all watched over a 48-hour period as roughly $285 billion in SaaS market capitalization evaporated. Thomson Reuters dropped 16% in a single day. The iShares software ETF entered a bear market. Traders at Jefferies coined a new term — the “SaaSpocalypse” — and within weeks, an estimated $2 trillion in software value had been wiped out.

The thesis seemed straightforward: if AI agents can autonomously handle the knowledge work that SaaS companies charge $30–$150 per seat per month to support, the per-seat model is structurally broken. Add the vibe coding movement — Y Combinator CEO Garry Tan reports that 25% of YC’s recent startups have codebases that are 95% AI-generated — and the fear is that companies will simply build their own tools rather than pay for SaaS.

The fear is real. The conclusion is wrong. And for investors in the Global South, the events that triggered the SaaSpocalypse are the best things that have happened to enterprise software since the advent of cloud computing.

The Death of SaaS Is Greatly Exaggerated

Let’s start with what the market is actually telling us. Software stocks have been hammered — the S&P North American Software Index traded below 20x forward earnings for the first time ever. But the earnings themselves? Still growing. Salesforce reported $41.5 billion in FY2026 revenue, up 10% year-over-year, with remaining performance obligations exceeding $72 billion. Snowflake grew revenue 30%. The disconnect between strong fundamentals and collapsing valuations is why many respected voices, including humble me, believe the selloff has gone too far. There will be price pressure over time and some companies will not create enough new value to compensate for lost margin, but leading SaaS companies can still evolve and grow.

Goldman Sachs CEO David Solomon called the selloff “too broad,” noting that “there’ll be winners and losers — plenty of companies will pivot and do just fine.” JPMorgan’s strategists described the market as pricing in “worst-case AI disruption scenarios that are unlikely to materialise,” pointing to “broken logic” in investors simultaneously believing AI will destroy all software while its capex fails to deliver returns. Wedbush’s Dan Ives called the selloff “a generational opportunity,” and Bill Gurley advised believers to “pick them up off the floor.”

I’m not a technical market analyst, and I won’t pretend to time the bottom. Many agree that the selloff was pricing in too much destruction, too quickly. But I believe the non-SaaS-insider bulls are closer to right than what the still-tanked market tells us today. Why? Outside the software industry itself, building and maintaining enterprise-grade software to address mission-critical business needs is not and will not become the core competency of any company. The idea that a logistics firm or a hospital chain will vibe code its own CRM or compliance system is, frankly, absurd. As Zoho CEO Sridhar Vembu pointedly asked: “Why don’t we see vibe-coded email or spreadsheet or accounting apps yet?”

That said, enterprise-grade agents are genuinely emerging. Agentic workflows combined with existing SaaS products can help companies move more quickly and at lower costs than ever before. The shift from per-seat to usage-based and outcome-based pricing is real — Gartner projects that by 2030, at least 40% of enterprise SaaS spend will shift to usage-, agent-, or outcome-based models. This is evolution, not extinction.

The Market Evidence: A US Valuation Problem, Not a Global Software Problem

Look at where the damage actually landed, and the pattern is clear. The SaaSpocalypse was fundamentally a repricing of overvalued US per-seat models — companies elsewhere that were never at those valuations didn’t have as far to fall.

In the US, the Goldman Sachs software basket crashed from roughly 34x forward earnings to below 20x. In Europe, companies that had reached US-style multiples fell in sympathy: SAP had its biggest daily loss since 2020, and Dassault Systèmes suffered its worst trading day in history (−21%) after reporting just 1% revenue growth. The sell-off punished any company trading at elevated multiples on the assumption of perpetual seat expansion.

India’s Nifty IT Index was hit hard — down 19% in February alone, its worst month since 2008 — but critically, this was driven by IT services companies (TCS, Infosys, Wipro) whose revenue comes from serving US and European enterprises. ICICIdirect estimated AI could dent 10–12% of their application development and maintenance revenue over 3–4 years. India’s domestic SaaS product companies — with Zoho as the shining star — are overwhelmingly private and serve global as well as local markets at more affordable price points. No SaaSpocalypse-specific crash.

In Brazil, Locaweb (B3: LWSA3) — the closest listed pure SaaS play — didn’t show a sharp SaaSpocalypse-driven decline; its multi-year slide predates AI disruption fears and is rooted in Brazil’s interest rate environment. TOTVS, Brazil’s largest ERP company, was absent from SaaSpocalypse coverage entirely. Across Africa and Southeast Asia, there was essentially no listed-market contagion — most SaaS startups are private, early-stage, and priced for local economies. As KPMG’s Marshal Luusa put it at the India AI Summit: “In Africa, AI must pay for itself early, or it doesn’t survive.”

The takeaway: the SaaSpocalypse is a correction for companies plugged into the US valuation ecosystem. Domestic SaaS companies in the Global South, serving local markets at local price points and local-currency economics, were largely insulated because they were never riding that bubble.

Why the Global South Should Be Paying Attention

Global SaaS was always priced out of reach. American and European SaaS tools are largely unaffordable for Global South companies beyond the largest multinationals. The cost structure of domestic enterprises in India, Southeast Asia, Africa, and Latin America has never supported $30–$150/seat/month costs. For most of these companies, the pre-SaaSpocalypse world was not one of SaaS abundance — it was a world of limited or no enterprise-wide software at all.

Agentic pricing changes the equation. Usage-based and outcome-based pricing models now allow direct correlation between AI token fees and business outcomes. A company pays for what the agent does, not for how many humans log in. As the Deloitte report on SaaS and AI agents notes, this shift makes enterprise software genuinely accessible for the first time.

Build-plus-buy becomes viable. The real opportunity isn’t “build vs. buy” — it’s “buy the platform, customize with agents.” Enterprises can adopt affordable cloud-based SaaS for systems of record, then vibe code point solutions for decision support and other areas aligned with localized needs that global SaaS would never address. Small models running on-premises and as edge AI on phones will reduce dependencies further. Open-source models — whose adoption is surging across Africa and the Global South — reduce reliance on any single provider.

Vertical SaaS with local moats is not exposed. The fastest SaaS growth globally is coming from largely un-digitized industries like healthcare, agriculture, and construction — precisely the verticals where Global South startups operate with local regulatory moats. As the cost of creating locally optimized SaaS falls, new opportunities emerge that were simply not economic before. The largely undigitized enterprises across the Global South will see options for affordable, locally-tuned digitization that did not exist before.

The Investment View

If I were a public equities buyer, I would buy historically low-priced SaaS companies that I believe have what it takes to evolve their models to agentic and usage-based pricing before competitive threats arrive. Deutsche Bank says the SaaSpocalypse has peaked, noting that it has not found “a single software company that expects a negative revenue effect from AI in 2026.” Software trading at a discount to the S&P 500 for the first time in modern history is either a historic buying opportunity or a permanent structural break. The companies still growing revenue at double digits, with $50–70 billion in contracted obligations, suggest the former.

As a Global South and India-focused VC investor in vertical AI, I see nothing but upside from the reduced costs, faster development times, and easier customization that agentic platforms provide. The SaaSpocalypse is a correction for overvalued American software giants. For the billions of people and millions of companies in emerging markets who were never served by those giants in the first place, it’s the beginning of something far more important: affordable, locally relevant, enterprise-grade software is becoming available, at population scale.

The real question today is not whether SaaS is dying. It’s whether AI will finally make it accessible to the rest of the world. I believe it will.

Latin America’s next generation of startups will be different from the last.

The first wave of venture-backed innovation in the region focused on bringing consumers online: digital banking, e-commerce marketplaces, and mobility platforms that connected fragmented demand and supply. Nubank, Mercado Libre, and Rappi defined this era.

But beneath that surface transformation, much of the region’s economic infrastructure remains unchanged. Old-school business practices still dominate in many areas, and ambitious businesses struggle to move money efficiently across borders. Supply chains remain fragmented, and industries that form the backbone of the economy continue to operate without the benefits readily available via modern digitization.

Today, a new wave of founders is beginning to address these structural inefficiencies — using AI to move us forward at higher speeds than ever before. The result may be a shift from building digital products on top of the economy to rebuilding the infrastructure of Latin America’s economy itself.

Three themes increasingly appear at the center of this transformation: low-friction financial transactions, AI productivity for already-digitized businesses, and AI-enabled transformation of offline businesses.

Reducing financial friction in a fragmented region

Moving and managing money in Latin America is a persistent point of friction.

Cross-border payments remain slow and expensive. Companies operating across multiple countries often juggle several banking relationships just to manage basic treasury operations. And access to capital remains uneven and unpredictable for small and medium-sized businesses. These friction points are precisely where some of the most interesting fintech innovations are happening.

Stablecoins, used by global players like Circle and startups like Avenia, use crypto technology to digitize major currencies, including the US dollar and Brazilian Real. They are not a speculative instrument like Bitcoin, but now offer institutional-grade settlement infrastructure. For businesses that need to move money between countries, the ability to settle transactions in minutes rather than days, and at a fraction of the current cost, fundamentally changes treasury dynamics. Lower transaction costs for B2B cross-border payments reduce financial friction and increase competitive edge for MSMEs; at the same time, programmable payments create opportunities for new financial workflows.

In parallel, lending is being reimagined through the platforms where economic activity actually happens.

For example, goFlux, a Brazil-based logistics-tech company, addresses this by building software that digitizes logistics workflows and adds fintech operations on top. Essentially, what started as a logistics management platform has evolved into a financial operating system for a freight ecosystem that’s fundamental to the entire economy. With real-time visibility into freight transactions across its network, the data becomes the foundation for financial services, enabling working capital financing to carriers secured by receivables.

In effect, for companies like goFlux, digitizing old-school business practices using modern software creates the data infrastructure for new financial services. Innovative financial services deployed in local, regional, or transnational models are becoming increasingly inseparable from the operational systems businesses use every day.

AI as a productivity layer for modern businesses

A second theme emerging across the region is the application of artificial intelligence within modern, already digitized business workflows.

For many years, enterprise software adoption in Latin America focused primarily on compliance and basic enterprise productivity. Tools helped automate accounting, payroll, regulatory, and sales operations processes, but rarely played a direct role in generating substantial revenue upside.

AI is beginning to change that equation. Startups like Diio and Dapta are building systems that sit closer to the growth engine of a company: identifying sales opportunities, qualifying leads, optimizing pricing, and automating parts of the customer acquisition process. By lowering the cost of accessing expertise and analysis, AI delivers capabilities that previously required expensive, specialized teams.

The impact can be particularly meaningful in markets where businesses operate with limited resources and fragmented customer access. When AI platforms are employed to help companies not only operate more efficiently but also generate more revenue, AI adoption accelerates with a positive feedback loop.

In this sense, AI expanding beyond productivity enables even small companies to gain economic leverage.

AI for the industries that still run on spreadsheets

Perhaps the most consequential opportunity lies in sectors that historically received little attention from startup founders and their venture investors.

Large parts of Latin America’s economy still operate through analog workflows with limited digitization. Manufacturing plants track production manually. Construction companies coordinate projects across disconnected tools. Logistics networks depend on phone calls and spreadsheets to manage complex operations.

At the same time, global supply chains are undergoing a structural shift.

Nearshoring is gradually positioning Mexico and parts of Latin America — including Colombia, Guatemala, and Costa Rica — as strategic production hubs for North American markets. The region’s proximity and labor base offer advantages, but capturing this opportunity requires meaningful improvements in productivity.

AI-native software platforms are beginning to emerge at this intersection. Some, like Pulsar and Prima, focus on optimizing industrial processes such as production planning, quality control, and predictive maintenance. Others, like Makasi and Solvento, digitize operational workflows across industries like construction, logistics, and field services. These systems often begin as workflow tools but gradually evolve into the system of record for daily operations. AI is enabling founders with deep domain expertise to create companies that solve “vertical” problems efficiently, quickly and profitably from anywhere, without requiring startups to be located in traditional tech hubs.

Once embedded deeply, vertical AI solutions become durable pieces of infrastructure, delivering profits to all who create and employ them.

A deeper phase of technological adoption

Taken together, these three development areas suggest a step-change in Latin America’s technology ecosystem.

The first generation of startups digitized consumer experiences and expanded access to financial services. That transformation was significant, but it largely operated at the surface of the economy.

The next generation will be addressing something deeper: the structural inefficiencies that shape how businesses operate, how capital flows, and how industries produce goods.Financial systems are being modernized. Operational workflows are being digitized. AI solutions are being applied to sectors that historically had very little software.

In a region where much of the economy is still analog, the companies that successfully introduce new infrastructure can become deeply embedded in how entire economies function.

And when that happens, the opportunity for long-term value creation will be far larger than any breakthrough since the mass introduction of mobile phones.

Will Poole is Co-Founder & Managing Partner of Capria Ventures, an applied AI investor with expertise in emerging markets.

AI directly addresses two of India’s oldest structural constraints — uneven access to expertise and uneven productivity across regions. As AI lowers the cost of accessing knowledge to close to zero and allows high-quality innovation to bloom faster and closer to the point of need, the economic value could be unmeasurably large.

This is why AI for Bharat is no longer just a policy aspiration. The Summit signalled this shift with large infrastructure announcements such as Reliance’s ₹10 lakh crore AI and green energy investment, and Google’s $15B India AI hub and skilling push. The parallel emphasis on DPIs, governance, and the MANAV framing of human-centric AI are promising signs for India’s AI leadership.

The AI genie inside industries

On the ground, AI for Bharat is becoming a real commercial opportunity — one already visible inside operating companies, small and large. For example, in operationally heavy businesses, friction often comes from routine process management of repetitive actions. BetterPlace, a frontline workforce operations platform, helps their customers process millions of resumes, compliance documents, and task records each month. Historically, their customers’ teams spent hours searching files, answering routine queries, manually reviewing applications, and onboarding workers who then churn away after only 3-6 months. After deploying AI systems that can instantly query documents and auto-evaluate resumes, entire layers of manual work disappeared. Tasks that once required large teams and long turnaround times now take minutes, allowing customers to operate with speed and clarity that would otherwise be unattainable.

Healthcare shows a similar pattern. At 5C Network, a diagnostics platform serving 2,000+ hospitals across tier-1/2/3 India, AI is embedded directly into radiologists’ workflows. Radiologists often review hundreds of scans per shift, where fatigue affects speed and consistency. Trained on 20+ million cases over six years — far beyond the ~90,000 scans a typical radiologist sees in a career — the system now flags abnormalities and drafts reports in minutes. This improves turnaround time and reduces error risk, freeing doctors to focus where judgment matters most. The Summit placed strong emphasis on this kind of “operationalization” of AI in healthcare, integrating AI into hospital architecture to address public health challenges.

These examples illustrate that AI’s immediate value in India will come from raising the performance ceiling of existing companies serving diverse industries. That is why education and talent development now matter as much or more than technology deployment. As India’s education reforms increasingly emphasize digital infrastructure, multidisciplinary learning, and practical skills, embedding AI literacy across universities and professional programs could prove more important than building new Desi-specific foundation models. When more people understand how to use AI fluently — most don’t need to know how to code with it — the productivity effects spread far beyond the tech sector.

The Multiplier Effect: Diffusion into Bharat

The mass diffusion effect is what makes AI uniquely suited to India’s economic structure. India’s linguistic, cultural, and market plurality means that companies that successfully scale here often have the operational resilience to expand to other complex markets.

For Bharat, the opportunity is innovation-driven productivity gains accruing across sectors. In agriculture, conversational models in local languages can translate research into field-level decisions. In education and healthcare, AI can extend scarce expertise across millions. In supply chains and manufacturing, it can improve quality and reduce friction, making distributed production more competitive. And when combined with fintech, AI enables faster risk assessment and the delivery of viable services to users who were previously too small to serve. This diffusion is already underway, and as AI literacy deepens, adoption will accelerate.

Encouragingly, India’s experience with telecom and DPI suggests that when access expands widely, the benefits tend to compound at population scale. As AI grows to reduce the cost of India’s structural constraints, a new generation of Bharat-first businesses is becoming not just possible — but investable.

Vision to Execution: India’s real AI challenge

However, none of this eliminates risk. AI-driven misinformation is already testing trust in institutions, and uneven adoption could create new disparities. The Summit also exposed execution gaps: a widely discussed incident involving a misrepresented “AI robot” served as a sobering reminder that scaling AI credibility requires deeper domestic R&D, stronger technical vetting, and institutional discipline.

I’m particularly excited about the MANAV initiative addressing labeling (marking content created/modified by AI) and authenticity (verification of content) to help preserve trust. India needs more than a clear vision and strong intent to realize its AI potential — it needs scaled execution hand-in-hand with practical regulation.

As PM Modi commented, “The real question today is not what artificial intelligence can do in the future; the question is what do we do with artificial intelligence in the present…”

Three years after ChatGPT’s public debut, the narrative around large language models (LLMs) feels very different. The story investors told themselves about capex ($350 billion in 2025!), enterprise adoption (actually falling), and Artificial General Intelligence (AGI) (nowhere in sight) is proving problematic. At the same time, concerns about job-evaporation have been somewhat mitigated, and the promise of generative and agentic AI as a useful tool in vertical markets, as well as general-purpose horizontal areas, continues to unfold.

With three years of rapid innovation behind us, it’s clear that AI’s practical gains are arriving unevenly. Meanwhile, the mass media and industry leaders’ narratives around jobs, infrastructure spending, and AGI are drifting farther from reality.

1. Jobs: Slower Shift, but with Tail Risk

While early reports show no immediate disruption to the job market — macro conditions, rather than AI and robots, are currently affecting white-collar jobs — anecdotal evidence suggests the shift is undeniably underway. Geoffrey Hinton, the “Godfather of AI,” recently warned that leaders haven’t yet considered the consequences of rapidly improving AI replacing cognitive labor in a demand-constrained economy.

For now, predictions about the speed of job displacement have proven wrong, but the probability remains high, with severity yet to be determined. AI looks more like a slow-burning general-purpose technology than a sudden labor apocalypse. Waiting for clear indicators, like rising unemployment, may leave us unprepared for a difficult challenge. It’s crucial to monitor early signs of employment slowdowns as AI’s influence on the workforce will unfold unpredictably, and unevenly, and maybe exponentially with major consequences for industry and society.

2. Capex-Revenue: Growing Disconnect

In 2025, Microsoft, Amazon, Alphabet, and Meta will have invested $344 billion in AI infrastructure, up from $228 billion in 2024. However, the revenue side looks stark. Annual AI revenue, not “someday” AI revenue, needs to hit $320-480 billion to support the capex build-out. This is a steep challenge, particularly as the vast majority of AI use is either through bundled offerings or is free to end users. OpenAI, for example, recently declared a “Code Red” due to competition from Gemini 3.

Despite the lack of ROI, adoption is surging. McKinsey research reveals that 88% of organizations use AI, but only 39% report EBIT impact, with most saying AI contributes under 5% to profits. Scaling beyond pilot phases remains tricky — a pattern Gartner’s Hype Cycle captured in 1995.

Adding to the uncertainty is rising algorithmic risk. DeepSeek’s release of version 3.2, which is 25–30 times cheaper than GPT-5 and Gemini, raises questions about the cost-effectiveness of today’s leading models. Yann LeCun, a deep learning pioneer and former head of AI for Meta, has critiqued LLMs as “not a path to human-level intelligence” and is now launching a new company focused on alternative AI architectures.

As respected experts pivot to new directions, there’s growing concern that a significant portion of today’s AI investments is misdirected. While giants like Microsoft and Google will likely weather the storm, others will struggle. Overbuilding, slow monetization, and rising uncertainty about AI’s future architecture suggest that a re-pricing of the sector — aka a bubble burst — is coming.

3. LLMs: Powerful but Not AGI

Gary Marcus, a prominent critic of LLMs, recently published a retrospective. He argues that AGI should no longer be the primary goal of AI research. Instead, he advocates for more targeted, verifiable tools and architectures that combine learning with explicit reasoning. This view is gaining traction. Yann LeCun’s critique of LLMs and push for world-model-based architectures aligns with Marcus, and even OpenAI co-founder Ilya Sutskever has acknowledged that the “just add GPUs” approach is over.

LLMs are growing and remain economically important. In enterprises, the future will likely center on small/medium models (SLMs) tailored to domain-specific needs. These models are more cost-effective, secure, and suited for demanding environments where 80-90% accuracy is a non-starter.

LLMs and SLMs remain increasingly powerful general-purpose technologies — but they are more akin to relational databases than AGI. This should temper expectations for AGI-driven explosive growth over the next 3–5 years.

2026: The Road Ahead

Taken together, these three hard truths point to a simple conclusion: AI will continue to dominate rhetoric and pilots, but will be under immense pressure to demonstrate economic value and to withstand sharper market scrutiny of revenue expectations. For those of us in the ring with this evolving technology, organizational agility and informed investments will continue to guide our way through 2026.

A version of this article was originally published on the Capria Ventures blog.

The AI boom has created extraordinary wealth — and mounting unease. In August, Jerry Kaplan, a Silicon Valley veteran who has experienced four tech bubbles, remarked, “We’re in the middle of an AI bubble. There is no need for dozens of companies to have foundation models.”

Similarly, other optimistic insiders now acknowledge that the market is overheated. The Bank of England recently warned that soaring valuations of tech giants like NVIDIA and Microsoft are reminiscent of the concentration levels seen at the peak of the dot-com bubble.

When this bubble does burst, it will unlikely be due to a single dramatic event, but rather to multiple intersecting factors, making it hard to speculate on its timing.

The Reasons We Haven’t Discussed —Yet!

While risks like the MIT study’s 95% AI pilot failure rate and tech giants’ “circular investing” have been widely noted, several other crucial factors haven’t received their due in mainstream commentary:

Trigger 1: The Game-Changing Efficiency Threat

If AI breakthroughs drastically reduce compute needs, billions in infrastructure investments could be rendered obsolete overnight. Take DeepSeek: in early 2025, the lab claimed to have trained GPT-4-level models for just $5.6 million — versus $78 million for GPT-4 — and to offer inference at a fraction of OpenAI’s cost. By September, DeepSeek reported another 50–75% reduction with its latest model.

Even if some claims are exaggerated, the direction is clear: algorithmic and tech innovation gains can collapse costs. This echoes the early-2000s telecom bust, when fiber-optic advances sent infrastructure costs plummeting.

Trigger 2: The Rise Of Task-Specific, Affordable AI

The second major pressure comes from buyers. Enterprises are realizing they don’t need — or can’t afford — massive, general-purpose models for many of their use cases. Gartner predicts that by 2027, companies will use small, task-specific models (SLMs) three times more than general-purpose LLMs. SLMs offer lower costs (up to 80% reduction), faster inference, and better data control. As adoption of SLMs increases, it undermines the revenue base that currently supports CapEx investments in large, generalized AI models.

Trigger 3: The Growing Gap Between Investment And Returns

AI infrastructure builders are spending heavily with little immediate revenue. In early October, Harris Kupperman estimated, based on 2025 CapEx projections, the industry would need $320–480 billion in annual revenue to break even. Currently, AI revenues are in the tens of billions. This CapEx-to-revenue gap is much larger than in previous infrastructure booms, like railroads (2×) or late-1990s telecom (4×). Once large-cap tech companies begin downgrading growth expectations, valuations could fall fast, exacerbating market instability.

Who Wins When the Bubble Pops

The collapse of the AI bubble won’t destroy all the value created — it will redistribute it, shaping winners and losers. In the aftermath, the groups best positioned to benefit are:

The Pragmatic AI Builders: SLMs and Domain-Specific Model Providers

Firms building efficient, task-optimized models and the infrastructure to manage them are best placed to capitalize on enterprise pragmatism. Providers with lower cost bases that focus on ROI — not brute-force scale — will align with the direction most businesses are heading.

The real-world AI integrators: Application & Service Innovators

As foundation models commoditize, value shifts to companies integrating AI into industry-specific workflows and business processes. Historically, this layer captures durable margins after infrastructure bubbles. To find winners, focus on companies behind the 5% of successful AI pilots in the MIT study, and solutions that boost the productivity of blue‑ and gray‑collar workers and SMBs.

The Resilient Infrastructure Titans: Hyperscalers with Staying Power

After the bubble, hyperscalers like Amazon Web Services, Microsoft Azure, and Google will consolidate overbuilt capacity. They may take short-term losses but will remain essential infrastructure providers in the AI economy — much like telecom carriers post-2001.

The Road Ahead for India

An AI bubble burst is no more the end of AI than the dot-com burst was the end of the Internet.

For India, the impact is likely to be less severe. Most AI investments are application-focused rather than in costly foundation models, meaning fewer losers and limited capital destruction. The investment in data centers (nearly 150 built already) will remain valuable, repurposed to meet domestic needs and support enterprise adoption across vital sectors like fintech, healthcare, logistics, etc. The most exposed will be conventional IT services firms and AI upstarts planning to drive revenue via sales to US businesses, which will slow down while overseas markets regroup post-bubble.

Turbulence is coming, and likely some of the biggest in living memory, and as individuals, companies, and nations, we would be well-advised to prepare for it.

A version of this article was originally published in The Economic Times.

I have been speaking to colleagues and friends whose college-age or recently graduated children don’t have a job or worry about getting one. In those discussions, I ask about how engaged their sons or daughters are with AI. Some don’t know. Others have provided surprising answers. For context: the people I’ve been speaking to are middle- to upper-level managers in large organizations, with plenty of resources, and clearly engaged in their children’s education.

One friend’s comments on his two early 20s-aged kids’ view on AI: “They hate AI.” I asked why. He said they feel it’s intellectually cheating and just don’t want to go there. Another friend with kids in high school and early college told me that his kids are not using it much unless required by school. He offered all three of them summer jobs with a good hourly wage if they would dive deep into an AI topic of their choice and produce something they are excited about by the end of the summer. Only one of three took him up on the offer and created an amazing AI-generated movie.

These conversations are just a few recent examples. Combined with the broader observations I wrote about recently, “AI Is Moving Too Fast and Too Slow”, I decided to create a pragmatic guide for what unemployed recent grads should be doing every day, and what students in their last year of school should start thinking about to prep for their upcoming job search.

Fundamentals first: your mindset matters most

Before getting to that pragmatic “what to do every day,” I want to revisit the key points Ravi and I have made in our book, Flourishing with AI. They revolve around mindset. Getting that right will make everything MUCH easier. Here’s a summary of Chapter 2 and Chapter 4 (which you can read in their entirety via our chatbot - see below) to get you thinking:

As a recent or upcoming graduate, you should embrace the evolving job market by reframing your thinking and focusing on possibilities rather than challenges. You should develop and/or hone your meta skills, which enable faster learning and adaptability, and cultivate an entrepreneurial mindset to create value rather than just seek stable jobs. Leveraging AI as a tool for experimentation, feedback, and efficiency can help you stand out in competitive environments. Graduates should also focus on blending technical skills with human-centric capabilities like storytelling, ethical thinking, and teamwork, referred to as “fusion skills.” Networking continuously to build authentic relationships and seeking mentorship can further enhance career opportunities. Finally, identifying areas where AI can complement your skills and focusing on tasks requiring unique human skills will help you get and stay competitive.

To which you might say “...yea, that’s all good, but I still don’t have a job - what do I do tomorrow?”

What to do every day

As I do with many things these days, I turned to AI as a thought partner to validate and build on my ideas, see which AI ideas I agree with and which I don’t, and then revise the combination to be something succinct. I asked five chatbots exactly the same question. Some answers were almost the same, some differed quite a bit. Here’s my synthesized and expanded answer, starting with the best of the five (from Gemini) for what to do every day:

Wake up and do these three things: Connect with One Person, Cultivate One Skill, Commit to One Application.

1/ Connect with One Person. Your network is your most powerful tool. Before you even look at a job posting, your first task is to reach out to one human being. This could be sending a personalized LinkedIn connection request to an alum from your college who works at a company you admire. (Do not just press “connect” -- adding a short personalized message will make the difference between accept and reject). You could also email or Link to a guest speaker who impressed you in a class to ask for a 15-minute “informational interview” or (I prefer “chat”) about their career path or about your evolved view on the topic they discussed. You can also follow up with a recruiter you spoke to at a career fair. Or a friend of your parents you met at a holiday party. (Many older folk are very willing to spend time helping their friends’ kids find their way in the world - don’t be shy to ask.)

The goal isn’t to ask for a job. The goal is to learn, make a connection, and get on someone’s radar for a future opportunity. This is how you find out about jobs that aren’t even posted yet.

2/ Cultivate One Skill. The skills expectation in today’s job market is moving faster than ever before. Spend at least 60 minutes every day making yourself a more valuable candidate. Ideas:

Read industry news to understand the current trends and challenges in your desired field. Since we’re talking AI and its incursion into almost every job, you should start with following AI thought leaders and newsletters. A few recommendations:

• The Neuron - https://www.theneurondaily.com/ - “550,000+ professionals from top companies like Microsoft, Apple, & Tesla and get the AI trends and tools you need to know to stay ahead of the curve.”

• TLDR AI - https://tldr.tech/ai “Get smarter about AI in 5 minutes. The most important AI, ML, and data science news in a free daily email.”

• Gary Marcus on Substack - https://garymarcus.substack.com/subscribe - Everyone’s favorite AI curmudgeon.

Watch a tutorial on a software program listed in job descriptions. Better yet, install the free version and use it to create something of interest, and update your blog with your impressions.

Work on a small personal project to add to your portfolio - publish your progress in your blog (and/or on Linkedin).

Refine your “elevator pitch” about yourself and/or a timely topic, and practice answers to common interview questions. I can’t overemphasize the importance of doing this regularly. If you’ve delivered your 60-second elevator pitch and your 3-minute cocktail pitch to your mirror, dog, friends, or whomever will listen 20 times, doing it the 21st time to someone who matters will be SO much easier.

Doing one or more of the above every day will keep you sharp, give you new things to add to your resume, and build confidence.

3/ Commit to One Application. Notice I said “one.” I’d rather you send one phenomenal, tailored application than ten generic, copy-pasted ones. There is an arms race with applicants using AI to write piles of untargeted applications and employers using AI to sift through them and reject most. Don’t get caught in that crossfire. Here’s how:

Read the job description carefully and customize your resume to highlight the exact skills they’re looking for. That does not mean a resume per job - but maybe you have a generic resume on LinkedIn, and you make 3-5 others that are for very different classes of jobs that you plan to apply to. You can easily create different versions on an AI-generated blog site using Lovable or Replit.

Write a compelling cover letter that tells a personal story and connects your experience directly to the company’s needs. Everyone is writing cover letters with AI. How do you make yours stand out?

Find the hiring manager on LinkedIn if you can and mention them in your letter. Or link to someone else in the company who might be able to connect you to the hiring manager.

This focused effort dramatically increases your chances of getting a response compared to just spamming the “Easy Apply” button.

That’s the guidance: Connect, Cultivate, Commit. The Rule of Threes.

If you do this every day, by the end of one week, you’ll have made 3-5 new professional contacts (if your outreach works), spent hours improving your skills, and submitted five high-quality applications. It creates momentum you can feel, turns a passive wait into an active process, and puts you squarely on the path to landing that first job.

A final point: allocate some time either on Friday afternoon or Monday morning to reflect on your efforts the previous week. Keep track of what you tried, what worked, what didn’t. Where did you keep to your program, where did you fall behind? What unexpected things did you learn that might tweak your plans the the following week?

And now that you have concrete things to do every week, please do go to https://jobbot.flourishingwith.ai/chat and head to Chapter 2 (which you can read or access via a podcast), summarized as follows:

Chapter 2 of ‘Flourishing with AI’ explores the critical need for an entrepreneurial mindset in the evolving job market. It emphasizes adaptability, meta skills, and leveraging AI for efficiency and creativity. The concept is further elaborated in the podcast ‘Flourishing Podcast- Ch 2- Entrepreneurial Mindset,’ which discusses the fading notion of stable, lifelong jobs due to AI-driven automation. It highlights the importance of adopting a ‘business of one’ mentality and developing adaptable capabilities, or ‘meta skills,’ to enable continuous learning, unlearning, and thriving amidst uncertainty.

Do you have other pragmatic suggestions for those looking to get a job or improve their career opportunities? Please send them my way, via LinkedIn https://linkedin.com/in/wpoole8

Recently, AI-driven platform Mercor, which recently raised $100 million in a Series B, was criticized for promoting gig work over traditional employment, raising concerns about job security and worker protections. While these concerns have some validity, they overlook a crucial reality: while unbridled application of AI and gig work could be tools of labor exploitation, these twin forces can be even more powerful enablers of economic opportunity, particularly in emerging markets where traditional employment models have failed millions.

AI as a Job Growth Engine

The assumption that AI reduces job stability ignores its role in unlocking new opportunities. AI-powered job platforms streamline recruitment, reduce hiring biases, and connect workers with employers faster than ever. The World Economic Forum’s 2023 Future of Jobs Report indicates that while automation may displace some jobs, it is also projected to create 69 million new roles by 2027.

Startups like Mercor exemplify how AI-driven recruitment can reshape the job market. In India, where over 1 million people join the workforce every month, AI-driven platforms like BetterPlace, Awign, Gigforce, and Naukri are improving job access and reducing hiring inefficiencies. In fact, for San Francisco-based Mercor, India remains the biggest talent source, followed by the US, with Europe and South America displaying rapid growth.

In regions where formal job markets are underdeveloped, traditional and AI-enabled platforms democratize access to work. A farmer in rural India, a coder in Nigeria, or a virtual assistant in the Philippines can now reach global clients based on the merits of their skills, without the gatekeeping of traditional employment structures. Companies like Fiverr, Upwork, and Andela have built billion-dollar businesses by facilitating remote work for skilled professionals in emerging economies. Similarly, companies like Wify and Apna are facilitating jobs for blue- and grey-collar workers, paving the way for more scale as AI-driven job matching and just-in-time training create even more opportunities.

The Gig Economy as a Pathway to Economic Inclusion

Critics argue that gig work erodes labor protections, but this perspective is largely rooted in developed-world employment models. In many parts of the Global South, the alternative to gig work is not a stable 9-to-5 job with benefits — it’s either unemployment or informal, unregulated labor with even fewer protections. Gig platforms bring workers into the formal economy, enabling them to earn, save, and invest in previously inaccessible ways.

India’s gig economy, for example, is projected to employ 23.5 million workers, making up 8-10% of the country’s non-farm workforce by 2029-30. AI-powered job-matching platforms are accelerating this growth by helping gig workers find opportunities tailored to their skills. Consider Anna Reddy, once a janitor who trained in furniture assembly through Wify. Within 10 months, his work and skill impressed his trainers, who promoted him to a junior trainer — his income was doubled and his career now has fresh potential. Stories like his are not outliers; they represent a global trend. The World Bank reports that 78% of gig workers in developing countries say they prefer platform-based work due to its flexibility and access to higher wages.

AI and Gig Work: A Call for Investment and Innovation

Instead of framing AI and gig work as threats, policymakers, businesses, and investors should collaborate on solutions that maximize their benefits. The gig economy is expected to grow at a CAGR of 17.4%, with countries in Asia, Africa, and Latin America witnessing a surge in gig work opportunities.

Investors have a unique opportunity to back ventures that provide AI-powered upskilling, fair wage guarantees, and payment security. Companies like Turing and Remotebase are already tackling these issues, proving that ethical gig work can be both profitable and sustainable.

Technology is not inherently oppressive or liberating — it is what we make of it. AI and gig work have already begun transforming lives in developing economies. The challenge is not to resist them but to shape them into engines of inclusive growth and invest in the infrastructure that makes them equitable and scalable.

A version of this article was originally published on the Capria Ventures blog.

My more in-depth thoughts on the hype cycle follow. I also have been posting short videos on relevant topics, which you can find here.

From the Peak of Expectations to the Trough of Disillusionment and Back

To understand where AI stands as of today, revisiting Gartner’s hype cycle is useful. This framework outlines the typical stages new technologies go through: from initial excitement and inflated expectations to disillusionment, and finally to a stable period of productivity. AI is currently navigating this familiar cycle, arguably with a mix of high expectations and challenges as it begins to find its place in everyday use.

Consider the media frenzy around ChatGPT 3.5, which sparked visions of AI-driven utopias and fears of mass unemployment. Now that the initial excitement has waned, reality is setting in. We’ve seen this pattern before with technologies, including the internet in the 1990s, solar power before ~10 years ago, and even railways back before most of us were born — overhyped initially, then undervalued, before eventually becoming mainstream. AI is now entering the “trough of disillusionment,” where expectations are being recalibrated again – this is at least the 3rd time, and there are more to come. But I believe the data show that its true potential is becoming clearer. This stage is natural for any advanced technology to mature and deliver sustained productivity and value.

While consumer-focused AI like ChatGPT sparked both excitement and hand-wringing doubts about the future, the real action is happening in enterprise use cases, where GenAI is quietly transforming productivity and business operations. Across sectors, companies are finding real-world applications for GenAI that go beyond the hype. Goldman Sachs, for instance, leverages its GS AI platform to handle over 100,000 daily transactions with data privacy and minimal errors. This example, along with Amazon’s cloud business being supercharged by GenAI, points to a broader trend of enterprise adoption taking hold, offering measurable ROI and meaningful digital transformation. Even massive consulting companies like BCG and McKinsey are aggressively adopting GenAI for internal use, signaling its real-world value.

Back to You, and to Our Startups

As the dust settles from the hype surrounding OpenAI, we’re seeing the emergence of no fewer than five comparable LLMs — deep foundation models from OpenAI, Google, Anthropic, Meta, and X.AI — that are shaping the next wave of AI-driven innovation in different ways. You’ve undoubtedly used ChatGPT, but I wouldn’t be surprised if you haven’t done more than sample one other. I find that’s frequently the case in discussions with busy people. I highly recommend giving each of them a decent trial period, as there is material differentiation depending on what you do. A great use case for investors or business executives is to treat the LLM as a “thought partner,” as described very well in Scott Galloway’s recent newsletter. The results will impress you.

As applied AI consumers of this tech, our startups are in a chaotic but very good place, as the foundation model giants are already fighting for market share, dropping API costs, extending credits, and adding multi-modal support, all while innovating in model performance and reasoning capability.

This broader enterprise adoption mirrors what we’re seeing in our own portfolio companies. For example, BetterPlace is already integrating GenAI into its services with its new training module generator for frontline workers. BetterPlace’s GenAI-driven module offers text-to-course generation in multiple local languages, now available to large enterprise customers such as Reliance and Swiggy, which collectively manage millions of frontline staff. This example illustrates that, despite the current skepticism surrounding AI, scaled startups are using the technology to deliver real-world solutions. No disillusionment here — just innovation and adoption where it matters most.

The Road Ahead: Ignore the Noise

So, where does all this leave us? As these examples show, the disillusionment phase doesn’t mean GenAI isn’t delivering value — it simply requires a longer-term perspective. Despite the media narrative, companies and investors who understand this are poised to reap the rewards.

From a mainstream perspective, we’re in the trough, and the skeptics are having their day. But if history is any guide, this is the time to pay attention. McKinsey’s research estimates that generative AI could add between $2.6 trillion and $4.4 trillion annually across 63 use cases.

ChatGPT now has a million business users, up from 600k three months ago. Capgemini’s recent study showed a 7.8% boost in productivity and a 6.7% increase in customer engagement, all thanks to AI. These productivity advances have not been seen since the heydays of enterprise PC deployment.

My take

GenAI’s enterprise value is becoming clear, even as we navigate the “trough of disillusionment.” The long-term trend is undeniable — incumbents and VCs alike are continuing to invest. Supporting this point, OpenAI co-founder Ilya Sutskever’s 3-month-old startup Safe Superintelligence raised $1 billion from prominent venture capital names in Silicon Valley, including Andreessen Horowitz and Sequoia Capital. Clearly, not everyone is wallowing in that trough — far from it. Those who recognize the underlying value are moving forward decisively.

The ElevenLabs deal shows the AI bubble is alive and well, with profit potential unclear due to the over-saturation of competing undifferentiated AI products and the assumption that incumbent tech giants are going to let startups have free reign in whatever segments they carve out. As seen in the August 30, 2023, WSJ article, savvy investors are increasingly concerned that young businesses have yet to prove they can retain users and develop products that existing tech companies and major infrastructure players couldn’t easily mimic.

In contrast, service-based companies in the Global South, including Southeast Asia, India, Africa, and Latin America, have started to capitalize on the opportunities that new AI technologies present. This is where Capria invests, passing over the eye-watering bubble valuations of the AI tools and infrastructure companies. We are seeing GenAI technologies advancing vital sectors including agriculture, healthcare, education, and fintech, all of which have been historically good businesses that now can grow faster and more profitably due to the application of the latest AI technologies.

AI Supporting Near-Term Economic Growth

Ignoring the “Skynet” naysayers for the moment, we are confident that AI will have a near-term positive impact on economic growth. Goldman Sachs Research estimates that widespread adoption of AI could contribute 1.5% to annual productivity growth over a ten-year period, lifting global GDP by nearly $7 trillion. The Global South’s young population and higher overall growth rates will disproportionately benefit from this prediction.

Recently, McKinsey & Company evaluated 63 GenAI use cases to solve specific problems across 16 business functions. The output was that AI could add $2.6 trillion to $4.4 trillion in economic benefits annually. Of that, 3.75% of the total annual value from GenAI use cases will be driven by customer relations, marketing and sales, software engineering, and R&D across all industry sectors. In combination with the expected increase in productivity alongside the technology, the total economic benefit would amount to an annual return of $6.1 to $7.9 trillion.

In India, productivity is expected to boost by 0.5%, with other countries including Mexico benefiting from a 0.6% productivity boost from 2022 to 2040. Productivity, which measures output relative to the input, or the value of goods and services produced divided by the amount of labor, capital, and other resources required to produce them, was the main engine of economic growth in the three decades from 1992 to 2022. This can be seen as illustrated above by a recent McKinsey report. With that, let’s take a closer look at how regions in the Global South are expected to perform in relation to their GDP growth and expected rise in productivity through the application of GenAI.

According to a recent EY report, India has the potential to add $359 to $438 billion to its GDP as a result of GenAI adoption. From 2023 to 2030, this would represent an additional 5.9% to 7.2% of GDP and an additional CAGR of .9% to 1.1%. What aspect of GenAI drives this growth?

Given that productivity growth makes up roughly 72% of GDP growth, the application of GenAI to spearhead India’s GDP is closely related to its impact on the country’s overall productivity. Of the many reasons for investors to be optimistic about the application of GenAI in developing countries, the largest is its impact on fostering intellectual growth alongside that of productivity rather than elimination of jobs.

Increase in Productivity Equals More Skilled Workers

In developing countries, the shortage of skilled workers is significant, and the application of GenAI can help reduce this gap by boosting productivity, thereby improving overall health and education levels that many of these countries struggle to attain.

Furthermore, the IMF predicts that labor markets in poorer countries will be less disrupted than those in wealthier ones. As a result of reduced employment disruption and increased productivity, we can expect the market in developing countries to significantly benefit from upgrades to operating models and the introduction of new products, while also revolutionizing economically dependent value chains.

LatAm’s largest sectors to see enormous benefits through applied GenAI

Along with Southeast Asia, LatAm is well-positioned for substantial economic growth driven by advances in AI. In a recent report by the Economist, countries including Argentina, Brazil, Chile, Colombia, and Mexico are thriving due to their private sectors leading AI development initiatives. Each of the region’s largest sectors has developed its own AI startup hubs, with initial regional success to back them up.

The largest sectors benefiting in the region are healthcare, Agtech, and Fintech. We see this firsthand with Kueski, the leading BNPL and digital payments player in Mexico and one of our portfolio companies, which uses GenAI to bring financial product education to a large market of first-time borrowers.

LatAm’s gold mine may well be agriculture, which provides the region’s largest volume of global exports. As seen in a report by the Economist, the sector provides fertile ground for growth, driven by large local markets and deep expertise. The Economist also highlights that the number of Agtech startups has increased by more than 600% between 2008 and 2015.

The addition of GenAI can have a significant impact in areas such as customer care and product development for the rising Agtech startup sector. Our portfolio company, Agrofy, has become the leading marketplace in Latin America for agriculture by simplifying transactions for merchants and farmers through the entire procurement process. For Agrofy, using GenAI will improve Q&A interactions with its customers, increase efficiency, enhance listing quality, and develop better product recommendations.

Global South startups have even more to gain from the use of AI

Compared to Western economies, those in the Global South are also able to take advantage of additional benefits of applied GenAI to tap into an expansive amount of market potential which has been previously hindered due to a lack of resources. Examples of this can be seen through the language “equalizer” which GenAI can provide.

GenAI provides local and global startups the opportunity to open up the vast markets of non-English speakers. The latest GenAI tech, such as Google’s Gemini, can already fluently translate between over 100 languages with its conversational model. Low-cost, high-quality translation into even lesser-spoken dialects will grow markets and decrease costs at the same time, including:

• Significant reductions in costs associated with human support teams as the business continues to grow through the application of GenAI chatbots, which can handle a wide range of queries and offer immediate and accurate responses in local dialects.

• Reductions in marketing costs via the time and resources spent on content creation and SEO optimization, ensuring content is not just engaging but also ranks well in local language search results.

• Enhancement of data analysis of large volumes of multi-lingual data, accompanied by summaries, trends, and insights, which are crucial for strategic decision-making.

• Automation and optimization of workflow productivity, by automating routine tasks such as email responses, scheduling, and document management, freeing up skilled employees’ time for more complex tasks.

My Takeaways

Last year was the year of awakening and experimentation. Some startups have already adopted GenAI and are benefiting from reduced costs and faster execution. But most are still in their early innings, learning what works and what doesn’t, evaluating the fast-moving new offers from providers big and small, and getting ready for scaled deployments in 2024.

Last year, we advised our startups to avoid getting locked into any one tech and to maintain their teams’ ability to devote the time needed to stay conversant with the latest tech, which is moving even faster than anyone expected. 2023 didn’t disappoint on the innovation front, and early signals are that 2024 will be even more exciting.

My colleagues and I feel privileged to be investing and supporting startups through these incredibly exciting times. If you’d like to invest alongside or with us, please drop me a note.

Let’s dispense with “worries” about Generative AI (GAI). The genies are out of the bottles and procreating in the clouds. They will infiltrate every aspect of society, quickly, for both good and bad. India is in a pivotal position. Again. The last existential global tech crisis started a quarter century ago as companies prepared for Y2K. Over-hyped unknowns of Y2K led to “worries” about hospitals shutting down and worse. But that didn’t happen, in part because the enterprising founders of Infosys, Wipro, TCS, and other then-startups but now IT giants partnered with customers worldwide to fix critical systems. 2023/24 will be India’s second Y2K moment, but bigger, as GAI will impact the very jobs the Y2K crises birthed as well as most other white-collar job in India, all at the same time.

India has more to gain and lose than any other country across the Global South (India, SE Asia, LatAm and Africa), particularly in the areas of IT and call center / BPM operations which employ about 5 million people, some 0.37 percent of India’s population. Call centers employ around 1.3 million Filipinos, about 1.2% of the population -- more to lose of a smaller population. Globally, the best leaders of call centers are working 24x7 to embrace GAI, retrain their agents to be escalation points, teach entry-level “prompt engineering”, and more. Notwithstanding aggressive adoption of GAI and retraining, a material percentage of front-line call center agents will be out of a job within a few years, everywhere. As will low-skill software testing engineers and less-skilled developers. A farmer advisory startup replaced 30% of call center staff with ChatGPT. If it works for India’s farmers already, it will go far.

GAI-driven job displacement is where policy needs to be developed - voluntary or regulated. Companies that employ soon-to-be-displaced workers have profited handsomely from the efforts of those workers. Their margins will improve as efficiency increases due to software-driven solutions. Before regulators impose it, CEOs could pledge to contribute a minimum of a year’s salary of each displaced worker to retain them. That’s a small price to pay to reduce the upheaval ahead.

There is no reason India should not continue to dominate the global IT services industry, and continue as an incubator and scaler of global-ready tech startups. When developer productivity goes up by 3x via GAI, think about how much more efficiently software will be produced everywhere. Entry-level developers will benefit the most from GAI which can help them both write and test code, increasing productivity and quality at the same time. If you manage developers, get them using GitHub CoPilot & ChatGPT quickly. If you manage in the BPM sector, get everyone trained on prompt engineering ASAP. And everyone must develop a data secturity policy so your data is not flowing to the GAI clouds. That will in turn bring efficiency to all, from VC startups in Bangalore to the 3+ lakh employees of Infosys nationwide.

But what will all of these highly productive developers do? IBM’s recent statement that it needs 30% fewer employees is a wake-up call. The answer lies in the digital revolution which has been underway for decades and yet has scratched the surface of digitizing the products and services that touch the four billion-strong Global South population. There is no reason that Indian IT services giants and innovative startups can’t set their AI-optimized sights on this digital revolution. Lower costs will make it practical to digitize the economy and employ lower-skill workers in new areas. India’s IT sector is already highly attractive to domestic and global investors. It only gets better.

(This article was originally published in The Economic Times)