An OpenAI model hacked into HuggingFace during a security evaluation. The French company reported that it used a Chinese open-source model in the detection and remediation of the attack. Global AI equity funding fell to 150 billion dollars in Q2 2026, down 37 per cent from Q1. Yet AI minted 37 new unicorns in Q2, up from 32 in Q1 and the strongest showing since Q2 2022. The global AI unicorn count now stands at 671.

Let's dive in.

Subscribe now

🗞️ The AI control debate is over. And China already won.

📢 What's the fuss all about?

The AI safety debate just became academic. An OpenAI model hacked into HuggingFace during a security evaluation, severe enough that it was reported to authorities before either company understood what was happening. The French company reported that it used a Chinese open-source model in the detection and remediation of the attack.

The same week, Moonshot AI released Kimi K3, a 2.8 trillion-parameter open-weight model that ranks number one on the Frontend Code Arena benchmark with a score of 1,679, ahead of Claude Fable 5 (1,631) and GPT-5.6 Sol (1,599). On Aikido Security's private cybersecurity benchmark, Kimi K3 found 23 of 26 recent vulnerabilities, matching OpenAI's GPT-5.6 Terra performance while being substantially cheaper.

In a leaked four-hour call with investors, DeepSeek founder Liang Wenfeng argued that China's only meaningful AI gap with the United States is compute, and that open source, low prices and team stability offer the best path to AGI. The US Treasury Secretary, Scott Bessent, responded by warning that the United States may consider sanctions and Entity List designations if open-source releases enable alleged PRC "covert, industrial-scale distillation attacks."

The irony? HuggingFace used an open-source model to address the cyber attack, not a closed one. Open source can democratise defending, not only widespread attacks.

💭 The Artificial Investor's thoughts

The moment Kimi K3 went open-weight at GPT-5 performance levels, the AI control debate became academic.

You cannot un-release a 2.8 trillion-parameter model. No amount of export controls, sanctions or Entity List designations can contain information that is already distributed globally. The genie is out of the bottle.

China optimised for inevitability while the West optimised for control. The HuggingFace breach exposes the core delusion: US labs spent billions on safety theatre — alignment research, red-teaming, responsible AI frameworks — while China focused on unstoppable distribution and cost efficiency. Every closed-model guardrail becomes irrelevant when open distillation works. Liang's leaked call makes this explicit: China's strategy is not to match US compute dollar-for-dollar, but to make compute constraints irrelevant through open-source distribution and aggressive cost reduction.

US Treasury's sanctions threat is a face-saving exercise. You cannot sanction open-source code that is already distributed globally. The policy response is 18 months too late. The containment strategy collapsed the moment DeepSeek proved that distillation works at scale and Moonshot demonstrated that open-weight models can match or exceed closed frontier models on critical benchmarks.

We see two counterpoints worth addressing. First, the United States still leads on some frontier model capabilities: GPT-5.6 Sol outperforms Kimi K3 on certain benchmarks, and closed models can iterate faster with proprietary data. Second, open-weight does not mean open training. China still faces compute constraints and US export controls on Nvidia's H100 and H200 chips limit their ability to train next-generation models at scale. But these advantages are temporary. The gap is closing and China's distribution strategy has already checkmated Western containment efforts.

The historical parallel is crypto export controls in the 1990s. The United States tried to restrict strong encryption, but once PGP was released as open source, the policy became unenforceable. Control fantasies collapsed. AI is information, not atoms. You can contain physical materials like uranium enrichment, but you cannot contain information.

Where does that take us? The real battleground becomes deployment and trust. Enterprises will choose models based on auditability, liability frameworks and integration ecosystems, not raw capability. Closed Western models retain an advantage only if they offer superior enterprise tooling, compliance infrastructure and legal indemnification. The technical race will be over at some point and get replaced by the governance race.

For founders and investors, opportunities lie in: 1) open-source infrastructure plays — tooling for deploying, fine-tuning and securing open-weight models will capture value as commoditisation accelerates, 2) enterprise trust and compliance layers — companies that solve liability, auditability and regulatory compliance for open models (model provenance tracking, usage monitoring, bias detection) will win as enterprises adopt open alternatives, and 3) vertical AI applications with proprietary data flywheels — the only sustainable moat in a world of commoditised foundation models is unique data that cannot be replicated through distillation.

🗞️ The great AI valuation reckoning begins

📢 What's the fuss all about?

Global AI equity funding fell to 150 billion dollars in Q2 2026, down 37 per cent from Q1's record 238 billion dollars, according to CB Insights' State of AI Q2'26 report. Yet AI minted 37 new unicorns in Q2, up from 32 in Q1 and the strongest showing since Q2 2022. The global AI unicorn count now stands at 671.

AI held its lead as the top exiting sector with 447 deals, but liquidity events are down: M&A fell 10 per cent and IPOs slipped 6 per cent. Meanwhile, AI agent traffic surged 7,851 per cent year-over-year, according to Cloudflare data. The companies behind the biggest exits were hiring aggressively beforehand, suggesting operational scale preceded liquidity.

💭 The Artificial Investor's thoughts

Funding down 37 per cent quarter-over-quarter while unicorn creation accelerates is a classic late-cycle signal. Early-stage rounds are still frothy (37 new unicorns in a single quarter), but growth-stage capital is repricing risk. This pattern preceded the 2022 tech correction. We're in the "still minting unicorns" phase, but the capital structure underneath is shifting.

Usage without monetisation is the new vanity metric. AI agent traffic up 7,851 per cent year-over-year sounds impressive, but M&A down 10 per cent means acquirers are not paying for usage. They are paying for revenue and margin, which most AI startups lack. The disconnect between usage growth and exit activity signals product-market fit without monetisation. This is a dangerous combination for late-stage valuations.

Our thematic analysis of the new AI unicorns reveals the hottest markets according to investors. Physical AI, robotics and embodied intelligence led the pack: Neura Robotics (Germany), Hark (US), General Intuition (US), Generalist AI (US), Bulage (China), Astribot (China), DexForce (China), Tianji (China), Odyssey (US), Standard Bots (US), Agilink (China), Botshare (China), and China's humanoid trio AI2 Robotics, PaXini Tech and Robotera.

Frontier and foundation model labs made up the second wave, spanning open-source and novel architectures: DeepSeek (China), Ineffable Intelligence (UK), Recursive (UK), Flourish (US), Sarvam AI (India), Mirendil (US) and ModelBest (China).

A picks-and-shovels layer of AI infrastructure and compute produced Ionic Digital (US), Armada (US), TensorWave (US), SiliconFlow (China), OpenRouter (US) and Runpod (US), while an emerging AI deployment and services layer (the businesses helping enterprises actually put AI to work) yielded OpenAI Deployment Company (US), the Anthropic Applied AI JV (US), Exa (US) and Blitzy (US).

Rounding out the field were vertical AI applications in healthcare, insurance, legal and security via Vi Labs (US), Corgi Insurance (US), Assort Health (US), Enter (Brazil) and Socket (US); AI for science, materials and generative media through CuspAI (UK), PhysicsX (UK), Evoken/Liblib (China) and Vast (China); and a heavily AI-driven defence and aerospace category anchored by Bezos-backed Project Prometheus (US) alongside Allen Control Systems (US), Mach Industries (US) and Twenty Technologies (US).

Our moat analysis of the new unicorns reveals structural weakness. Of 671 AI unicorns, we estimate only around 15 per cent (roughly 100 companies) have durable advantages: proprietary data flywheels (healthcare AI with longitudinal patient data), regulatory capture (financial compliance AI) or embedded workflows with high switching costs (coding assistants integrated into IDEs). The other 85 per cent face margin compression as foundation models commoditise. If your moat is "we fine-tuned GPT-5 for legal", that advantage evaporates when OpenAI releases GPT-5.7 with better legal reasoning out-of-the-box.

The historical parallel is the dot-com bubble. Funding peaked in Q1 2000, then collapsed 80 per cent over 18 months. Unicorn creation continued through mid-2000 before the reckoning hit. We are in the "still minting unicorns" phase. Cloud infrastructure in 2011 to 2012 offers another lesson: usage metrics (AWS revenue growth, GitHub stars) surged, but M&A and IPO activity lagged by two to three years. The market eventually caught up, but many early cloud startups got acquired at flat or down valuations.

We expect M&A activity to pick up in H2 2026 as acquihires accelerate. BigTech will buy AI talent and technology at discounted valuations rather than compete with overfunded startups. Expect 50+ AI acquihires in the 10 to 50 million dollar range. The 15 per cent with durable moats will capture 85 per cent of value creation. Vertical AI with proprietary data, regulatory moats or embedded workflows will trade at 15 to 25 times revenue multiples. The rest will compress to 3 to 5 times as they become feature companies.

🗞️ The world's first humanoid cage fight is a reality

What's the fuss all about?

The world's first humanoid cage fight showcased robotics capabilities in a staged entertainment event. Two humanoid robots, designed to mimic human movement and combat techniques, squared off in a controlled arena, generating significant media coverage and social media engagement. The event was positioned as a demonstration of advanced robotics and AI-driven real-time decision-making.

Promoters framed the cage fight as a milestone for humanoid robotics, arguing that it proves the technology is ready for consumer-facing applications. The spectacle drew comparisons to early automotive demonstrations and aviation stunts, with proponents claiming it will accelerate public interest and investment in humanoid platforms.

We will be taking a break for the summer. See you in a few weeks with our summer recap version. Enjoy the summer!

Share

Oracle posted what looked like the greatest earnings print in enterprise software history, only to watch its stock drop by 12% as investors realised the 638 billion dollar backlog was a leveraged bet on customers who are themselves burning cash at unsustainable rates.

Meanwhile, two contradictory narratives emerged on AI employment: Gartner projects net job creation by 2028, while nearly 200 economists warn governments to prepare for rapid displacement. Both are wrong about the timeline; we explain below why.

And in New York, a school district began piloting Sally, a seated humanoid robot for STEM tutoring.

Let's dive in.

Subscribe now

🗞️ Oracle's AI demand backlog masks a 24 billion dollar cash flow hole while OpenAI burns nearly 4 billion dollars per quarter

📢 What's the fuss all about?

Oracle reported a staggering 638 billion dollar backlog in Q3 2026 (up 363% year-on-year) driven by multi-year AI infrastructure contracts with hyperscalers and AI labs. On paper, this was the greatest earnings print in enterprise software history. The market's response? A 12% stock drop the next day.

The reason: Oracle posted negative 23.7 billion dollars in free cash flow during the same quarter, according to Reuters. The disconnect between reported backlog and actual cash generation raised immediate questions about whether Oracle's customers could afford to honour those commitments. OpenAI, one of Oracle's largest customers, burned 3.7 billion dollars in Q1 2026 against 5.7 billion dollars in revenue—a 65% burn rate that suggests the company is far from self-sustaining profitability.

The market correction extended beyond Oracle. S&P downgraded Oracle's credit rating to BBB- and short sellers built a 25 billion dollar bet against SpaceX as AI-adjacent stocks retreated. SpaceX shares fell below their IPO price, with short sellers banking 8.7 billion dollars in profit as the AI infrastructure trade unwound. Quick self pat on the back here, as we predicted that very soon there would be a lower price entry point for SpaceX—we were proven right within a month.

The broader correction, including the collapse of Fable 5 and other IPO retreats, signals growing market scepticism about the AI sector's path to sustainable profitability.

💭 The Artificial Investor's thoughts

This is the first serious stress test of AI infrastructure economics and the results are sobering.

The S&P downgrade and 12% stock drop signal the market sees Oracle as a leveraged bet on AI profitability, not a safe infrastructure play. SpaceX short profits and the broader AI stock correction suggest investors are re-pricing AI infrastructure risk across the board. We've seen this movie before.

In the dot-com era, telcos built out fibre infrastructure on the assumption that dot-coms would pay for bandwidth. When the dot-coms went bust, the telcos were left holding stranded assets—Global Crossing and WorldCom both filed for bankruptcy. Cloud in 2010 to 2014 was different: Amazon's AWS posted negative free cash flow for years while building data centres, but AWS customers (both enterprises and startups) had sustainable unit economics. Oracle's customers don't yet.

The solar industry in 2008 to 2012 offers another parallel. Panel manufacturers overbuilt capacity assuming demand would scale linearly, but when subsidies dried up and prices collapsed, the industry consolidated brutally—Solyndra and Q-Cells went bankrupt. Oracle's backlog may be a similar mirage: demand that looks real on paper but evaporates when customers run out of runway.

Where does that take us? We think Oracle's 363% backlog growth is a lagging indicator, not a leading one. The real question is whether OpenAI and other AI labs can reach profitability before their capital runs out. If they can't, Oracle's backlog will convert to write-offs, not revenue. The market is already pricing in that risk—hence the 12% stock drop and the S&P downgrade.

For founders and investors, the takeaway is sharp: 1) AI infrastructure is a leveraged bet on AI lab profitability, not a safe play, and 2) watch OpenAI's burn rate closely—if it doesn't materially improve by Q4 2026, expect a broader correction in AI infrastructure valuations.

🗞️ Gartner says AI creates jobs by 2028, but 200 economists warn displacement is coming. Who is wrong?

📢 What's the fuss all about?

Two contradictory narratives emerged last week on AI's impact on employment. Gartner projects that AI will create more jobs than it eliminates starting in 2028, driven by new roles in AI oversight, data curation and compliance. The consulting firm's forecast assumes enterprises will hire AI trainers, prompt engineers and compliance officers faster than they automate existing roles.

Meanwhile, nearly 200 economists and tech leaders signed an open letter warning governments to prepare for rapid AI job displacement, citing risks to low-skill and mid-skill workers. The letter reflects a shift from "gradual impact" to "rapid disruption" warnings, citing accelerating AI capability and falling deployment costs.

Early data offers mixed signals. Stanford research shows a 20% drop in young developer employment (ages 22 to 25) since 2024, suggesting early-career tech jobs are already being displaced. However, the World Economic Forum estimates 10% to 80% more new jobs created versus the ones displaced by 2030, though it acknowledges transition friction will be severe.

💭 The Artificial Investor's thoughts

Both forecasts are wrong about the timeline. We expect AI job displacement to properly arrive in 2030 to 2035, not 2026 to 2028, because regulatory friction, union resistance and risk-averse enterprise procurement are slowing down deployment velocity in the real world.

Gartner's 2028 timeline assumes frictionless adoption. But regulated industries—banking, healthcare, government—move at bureaucratic speed. Banks still run COBOL systems from the 1970s. Hospitals require five-year IT validation cycles. The idea that these sectors will deploy AI fast enough to displace workers by 2028 is fantasy. Enterprise IT departments are expanding AI teams to manage deployments, not replacing human staff yet.

The economists' warning letter will likely trigger further preemptive policy action. EU AI Act-style frameworks will likely spread globally, slowing deployment faster than tech optimists expect. Compliance kills velocity.

We've seen this before: industrial automation in the 1970s to 1990s was predicted to eliminate manufacturing jobs by 2000, but union resistance, safety regulations and high upfront costs delayed adoption by 15 to 20 years. Displacement happened, but much slower than forecast.

Offshoring in the 2000s offers another parallel. Consultants predicted rapid job losses to India and China, but language barriers, time-zone friction and quality-control issues meant the transition took a decade longer than expected. Self-checkout in the 2010s was supposed to eliminate cashier jobs, but theft, customer frustration and union pushback meant most stores kept human staff. Automation happened at the margins, not wholesale.

Stanford's 20% developer drop may be a Silicon Valley phenomenon, not a leading indicator. Junior developers in startups are being displaced because startups move fast and have no unions. But enterprise IT departments, where the bulk of developer jobs sit, are hiring AI specialists, not firing coders. The first wave of displacement will hit non-regulated, low-union sectors: customer service, content moderation, junior coding roles. The bulk of knowledge-work jobs—law, finance and healthcare—will be insulated until 2030 or later.

Here's the thing: governments will pass preemptive AI labour laws. Mandatory human-in-the-loop for high-stakes decisions. Retraining subsidies. AI deployment taxes. These will delay the reckoning but won't prevent it. The real displacement arrives 2030 to 2035, not 2026 to 2028.

For founders and investors, the takeaway is that AI compliance tooling is the near-term opportunity. 1) Build tools that help enterprises navigate AI labour regulations—audit trails, explainability dashboards, risk-scoring systems, 2) vertical applications in regulated industries (law, finance, healthcare) will face slower displacement and offer defensible moats, and 3) retraining platforms and upskilling infrastructure will see government-backed demand as displacement fears mount, even if the actual displacement is years away.

🗞️ New York schools pilot Sally the humanoid robot for STEM tutoring

A New York school district is piloting Sally, a seated humanoid robot designed to assist with STEM classes and provide round-the-clock homework support. According to Mashable, Sally is one of the first physical AI deployments in K-12 classrooms, not just a software chatbot. Students interact with a robot body, not a screen.

The pilot raises immediate questions about whether humanoid robots can replace or augment human teachers, and whether the form factor matters for learning outcomes. Early media coverage focuses on the "creepy" factor—whether students will have nightmares or accept the robot as a normal part of education. Sally is seated, not walking, which reduces cost and complexity while still providing a "person-like" interaction.

See you next week for more AI insights.

Share

My name is Aris Xenofontos and I am an investor at Seaya Ventures. This is the weekly version of the Artificial Investor that covers the top AI developments of the last days.

The AI industry's geopolitical and economic foundations are fracturing simultaneously. The US Commerce Department's decision to block foreign nationals from accessing Claude Fable 5, a model already in broad commercial use, marks the moment when AI dependency became an existential risk for every non-US country. Meanwhile, OpenAI's launch of its own AI chip signals that the industry believes current training architectures have plateaued, heating up the competitive battleground for Nvidia. And beneath both trends, a structural memory shortage through 2030 is creating a permanent two-tier AI economy: hyperscalers with chip allocations build strong moats, while everyone else becomes API renters with no path to vertical integration.

Let's dive in.

Subscribe now

🗞️ Sovereign AI just became inevitable, not optional

📢 What's the fuss all about?

The US government has crossed a critical threshold in AI export controls. According to The Information, Sam Altman told OpenAI staff that the government requested a staggered rollout of GPT-5.6, approving access "customer by customer" over security concerns. This marks the first time a commercially available AI model faces de facto export licensing.

The controls escalated further when the US Commerce Department ordered Anthropic to block foreign nationals from accessing Claude Mythos 5 and Fable 5, according to Global Government Forum. The restrictions affect H1-B visa holders working at US companies, international research institutions, and multinational corporations. Al Jazeera reports that companies like S&P Global and universities with foreign researchers now face operational disruptions as they lose access to models they were already using in production.

This is the first time export controls have been applied retroactively to models already deployed commercially, not experimental research prototypes or chips.

💭 The Artificial Investor's thoughts

This is the moment AI dependency became a geopolitical liability that no serious country can ignore.

The US government isn't blocking experimental military AI or dual-use research systems—it's blocking commercially available models that thousands of companies worldwide were already using for coding, customer service, and data analysis. S&P Global, research institutions, and foreign companies now understand they're one Commerce Department memo away from losing critical infrastructure. That's not a risk any sovereign government can accept.

We think this will trigger a surge in domestic AI investment across the EU, Japan, South Korea and India within 12 months. The political calculus just shifted: building indigenous frontier models was previously a "nice to have" for technological prestige. Now it's an existential necessity. Countries cannot build modern economies on infrastructure they don't control. AI is proving to be critical infrastructure like telecom or energy grids.

The historical parallel is clear. When the US banned Huawei from 5G networks in 2019-2020, the EU and India accelerated domestic telecom infrastructure programmes. The same dynamic is now playing out in AI, but with higher stakes. 5G was about network equipment; AI is about the cognitive layer of every digital service. Losing access doesn't just mean slower internet—it means your financial systems, healthcare diagnostics, and government services stop working.

The counterargument is that building frontier models requires 10 billion dollars and three to five years, which most countries lack the capital or talent pools to execute. That's true, but misses the point. First, China's DeepSeek success shows indigenous models are technically feasible—the missing piece was political will, which the US just provided. Second, the EU, Japan, and South Korea already have semiconductor expertise, cloud infrastructure and capital markets. What they lacked was the urgency to deploy those resources at scale. Third, open-weight models like Llama, Mistral, and Nemotron still exist—countries can fine-tune rather than train from scratch, which is far more economical.

We also don't think US controls will be temporary or narrowly scoped. The customer-by-customer approval regime for GPT-5.6 and the foreign-national block on Claude are structural, not tactical. They signal that the US views frontier AI as a strategic asset to be rationed, not a commodity to be traded. That calculus doesn't reverse when the next administration takes office—it only intensifies.

For founders and investors, the implications are sharp: 1) non-US AI infrastructure—compute, data centres, model hosting—becomes a strategic asset. Expect sovereign cloud providers in Europe and Asia to raise billions. 2) model training, fine-tuning, and deployment tooling will see explosive demand as countries avoid US dependency. 3) the AI market will fragment instead of consolidate. The fantasy of a single global AI platform is over. We're heading into a world of regional AI ecosystems, each with its own models, regulations and supply chains.

🗞️ The training-inference split marks AI's maturity—and Nvidia's vulnerability

📢 What's the fuss all about?

OpenAI announced the launch of its custom inference chip codenamed Jalapeño in partnership with Broadcom, according to Cloudflare's reporting, despite never having built hardware before.

A few days earlier, Google split its Tensor Processing Unit (TPU) architecture into separate training and inference chips. Google's blog post announces the TPU 8t for training and the TPU 8i for inference, marking a strategic shift in how hyperscalers allocate chip resources.

This comes after Meta shipped four generations of its MTIA (Meta Training and Inference Accelerator) chip in just two years, according to their announcement, all optimised for recommendation inference rather than training.

This validates that inference, not training, is the bottleneck for their business model.

💭 The Artificial Investor's thoughts

This is the signal that the industry believes current training architectures are "good enough" and that the next decade's competitive battleground is inference optimisation, not training capability.

Why does this matter? Agentic workloads—code generation, search, customer service—are inference-heavy. Every time a user queries ChatGPT or Claude, that's an inference call. Training runs are one-time capex; inference is ongoing opex that compounds with user growth. Optimising inference has 10 to 100 times ROI compared with training. That's why OpenAI, a company that has never designed chips, is now building custom inference silicon. They're betting that inference cost, not model capability, determines long-term margins.

The bifurcation also signals something deeper: hyperscalers are hedging against Nvidia's monopoly. By designing specialised inference chips, they reduce dependency on Nvidia's roadmap and pricing. Google's TPU 8i, Microsoft's Maia 200, and Meta's MTIA are all inference-optimised. None of them compete with Nvidia's H100 or Blackwell for training, but all of them reduce the need for Nvidia chips at inference time.

This is a threat to Nvidia, but not an existential one. Yet.

Nvidia's defensibility lies in three areas: software superiority (CUDA ecosystem), developer lock-in (every ML engineer learns on Nvidia), and cutting-edge technology (Blackwell, Vera Rubin). But inference is a different battlefield. Inference chips don't need the raw compute density of training chips—they need low latency, high throughput and energy efficiency. That's a game Google, Microsoft, and Meta can win, because they understand their own workloads better than Nvidia does.

The counterargument is that Nvidia's head start in software, supply chain and ecosystem is a moat that doesn't erode quickly. That's true, but it also means the window for Nvidia to respond is narrowing. If inference chips become the dominant revenue driver by 2027 to 2028, Nvidia's training-focused roadmap becomes less relevant.

For founders and investors, the takeaway is sharp: 1) inference infrastructure—chips, orchestration, caching—is the new margin battleground. Companies owning this layer capture more value than model providers. 2) training-focused AI companies (research labs, foundation model startups) face margin compression. Inference-optimised models like Cursor Composer and Mistral variants win on economics. 3) vertical integration into inference chips is now essential for hyperscalers. Ambitious AI scaleups without chip roadmaps will be API renters with no path to autonomy.

🗞️ Memory shortage creates the great AI stratification

📢 What's the fuss all about?

Apple raised Mac and iPad prices by 15 to 25 per cent due to soaring memory and storage chip costs, according to The Wall Street Journal, with executives saying, "We have never seen a component price increase this much, this quickly." The price hikes aren't limited to consumer electronics—ABC7 reports that the AI data-centre boom is driving chip shortages and inflationary pressure across the entire electronics supply chain.

Hyperscalers' 741 billion dollar AI spend is creating persistent inflation, per Business Model Analyst, as memory and storage demand outstrips production capacity. Fortune warns that rising chip costs threaten the AI economy, as production can't keep pace with demand. CNAS reports that AI chip supply constraints are pricing out researchers, with rental costs rising despite efficiency gains.

The memory shortage is expected to persist through 2030, according to Sourceability, as AI demand creates an allocation-only market where spot purchases are no longer possible. Even well-funded AI companies are supply-constrained: Anthropic implemented stricter rate limits, and Google Cloud admits inference capacity is "supply constrained."

💭 The Artificial Investor's thoughts

This is the moment the AI economy bifurcates into two permanent tiers: hyperscalers who own chip allocation and frontier AI labs who rent at uncompetitive rates.

The key insight is that memory markets have shifted from spot pricing to allocation-only. That means chips aren't for sale at any price—they're rationed to customers with long-term supplier relationships. Microsoft, Google, Amazon, and Meta control 80 per cent or more of global AI chip allocations because they signed multi-year contracts with Samsung, Micron, and SK Hynix years ago. Startups, researchers and SMBs are structurally locked out. They can't outbid hyperscalers, because there's no auction.

This creates a vicious cycle. Hyperscalers with chip allocations can build vertically integrated AI products (models, inference, applications) at marginal cost. Startups building on APIs (OpenAI, Anthropic) have no path to cost reduction; they're perpetually paying retail prices for compute, while hyperscalers pay wholesale. Over time, that margin gap becomes dangerous. The evidence is already visible. Anthropic rolled out stricter rate limits despite being backed by Amazon and Google.

The counterargument is that memory prices are cyclical and 2027 to 2028 may see oversupply as new fabs come online (Samsung, Micron expansions). That's plausible, but the timing matters. AI inference memory requirements are growing faster than DRAM and HBM production capacity. Even if new fabs come online, they'll be absorbed by hyperscaler demand before they reach the spot market. The allocation-only regime persists.

Another counterpoint: efficiency gains, such as quantisation and KV cache compression, could reduce memory requirements faster than demand grows. We've covered TurboQuant and KVTC in prior issues, and those techniques do deliver six to 20 times memory reductions. But efficiency gains don't help if you can't buy chips in the first place. The bottleneck is access, not cost per token.

A third counterpoint: cloud partnerships (Anthropic-AWS, OpenAI-Microsoft) give frontier AI labs indirect access to hyperscaler allocations. That's true, but it also means AI labs are structurally dependent on hyperscalers. They can't negotiate pricing, they can't control infrastructure roadmaps and they're one contract renegotiation away from margin compression. That's not a defensible position.

The historical parallel is the 2010s smartphone OLED shortage. Samsung's vertical integration (owning both display manufacturing and phone production) crushed competitors who couldn't secure screens. LG, HTC, and others either exited the market or became niche players. Non-hyperscaler frontier AI labs face the same risks.

For founders and investors, the implications are obvious: 1) AI startups without hyperscaler partnerships or chip allocations face structural disadvantage with margin improvement or cost competitiveness. 2) memory and storage companies (Micron, SK Hynix, Samsung) are the real AI infrastructure winners. Pricing power is locked in through 2030. 3) vertical integration is the only defensible strategy. Companies that don't own chips, data centres and models will become API resellers.

Our prediction: by end of 2027, a significantly larger share of AI inference runs on hyperscaler infrastructure (famous last words of an AI investor :-)). The memory shortage will trigger a wave of vertical integration. Companies like Databricks and Snowflake will acquire chip startups or sign five-year-plus supply agreements.

See you next week for more AI insights.

Share

Let's dive in.

SpaceX's closing price at 160 dollars per share represents the market's most audacious valuation call on AI infrastructure; institutional oversubscription at multiple times available shares signals sophisticated capital views Starlink as a strategic monopoly with an irreplaceable connectivity moat, while the 2 trillion dollar price tag requires heroic revenue growth and margin assumptions never achieved by capital-intensive businesses. The bull case rests on Starlink's pricing power and xAI synergies; the bear case warns that 2 trillion dollars on 5 billion dollar losses leaves no room for execution stumbles, competitive pressure or AI bubble correction. This deep-dive examines both sides and offers a tactical view for founders and investors navigating the most consequential infrastructure IPO since Amazon and Google.

🐂 The bull case: strategic value, exponential revenue growth, underlying profitable unit economics and Musk's execution track record provide a once-in-a-decade infrastructure stock holding opportunity

💲 Show me the money

Starlink alone provides a path to a more digestaile valuation of about 30x revenues by reaching 60 billion revenues. At the current revenue per user of about 800 dollars per year, the subscriber base would need to grow from the current levels of 10 million to 75 million subscribers. Taking into account that there are about 1.3 - 1.5 billion fixed broadband users globally, this implies a global market share of 5%.

And it's not just about revenues; profitability is important too. All three core business units provide a path to group profitability. Starlink is profitable and is expected to grow at least 7x, as mentioned earlier, which means that its positive profit contribution will have a higher weight at the group level. Starship currently loses money, but costs per launch are reducing every year and, if the cost curve continues, Starship is expected to reach its breakeven point in the next 5 years. Finally, xAI, the most loss-making business unit, should be able to generate profit assuming AI demand continues to grow and forces extraterrestrial distributed compute.

🏰 A real competitive moat

Starlink's connectivity moat enables xAI training at scale in ways terrestrial infrastructure cannot match. Low-latency global coverage allows distributed compute across remote data centres, data collection from sensors in oceans and deserts, and edge AI deployment that compounds with each satellite launch. As xAI scales its training clusters, Starlink becomes the connective tissue linking geographically dispersed GPUs; a capability no terrestrial provider can replicate.

The network effects are real. Every new satellite increases coverage density, reduces latency, and expands the addressable market for AI workloads that require global reach. This isn't speculative—it's operational reality. Starlink already powers remote oil rigs, shipping fleets, and military operations where terrestrial alternatives don't exist.

♟️ The intangible strategic value

Starlink strategic irreplaceability creates monopoly-like pricing power. It's the only credible Plan B for global communications; immune to terrestrial infrastructure failures, geopolitical disruptions, and natural disasters. Governments and enterprises will pay premium prices to access this capability regardless of near-term financials.

Militaries need Starlink for secure communications in conflict zones. Shipping fleets need it for real-time navigation and cargo tracking. Airlines need it for in-flight connectivity. Remote operations (mining, agriculture, energy) have zero alternatives. This inelastic demand supports forty per cent plus net margins at scale because the intangible strategic value exceeds financial metrics. When the alternative is "no connectivity," customers will pay whatever Starlink charges.

🪄 The "Musk factor"

Value potential is backed by execution track record. Musk scaled Tesla from near-bankruptcy to an eight hundred billion dollar market cap, and Starlink deployed twelve thousand plus satellites in seven years - faster than any government or competitor. Patient capital that bet on Tesla at fifty billion dollars in 2019 saw fifteen times returns despite years of losses. The lesson? Exceptional founders with track records of turning capital-intensive moonshots into dominant platforms deserve premium valuations.

⛳ The floor

Gulf sovereign wealth funds pouring billions into SpaceX orders adds a geopolitical dimension that typical IPOs lack. These funds operate on 30 to 50 year horizons and view Starlink as strategic infrastructure—not a quarterly earnings play. For Gulf states diversifying away from oil dependence, owning a stake in the world's only credible global satellite internet provider is a hedge against geopolitical fragmentation; it's a bet on AI infrastructure as the new oil. This sovereign capital creates a structural bid floor. In addition, at a valuation of 2 trillion dollars, SpaceX enters S&P 500 and MSCI indices, so passive funds must buy the stock. Taking into account the structural growing share of passive investment funds, this will likely create further market price stability.

🐻 The bear case: 2 trillion dollars valuation on 20 billion revenues and 5 billion losses represents peak AI bubble pricing that requires heroic assumptions and leaves no room for execution stumbles, competitive pressure or AI bubble correction

🧮 The math doesn't add up

Reaching 60 billion dollars of Starlink revenue would require roughly 75 million subscribers at today's pricing - about seven times the current 10.3 million and some 5% of all fixed-broadband connections on Earth - but the pricing won't hold. Blended ARPU has already fallen from about 99 (in 2023) to 66 dollars by Q1 2026, an 33% decline even as the subscriber base quadrupled, precisely because growth is now coming from price-sensitive markets across Africa and South-East Asia. This is structural: the broadband users still unserved are overwhelmingly at the bottom of the global income pyramid, where affordability is the dominant barrier and an entry-level device can cost 87% of monthly income for the poorest households. Every incremental subscriber is therefore poorer than the last, which means the very volume the bull case needs drives ARPU down further - the model fights itself.

Also, many investment bankers' models rely on a 40% net margin for the business. However, capital-intensive infrastructure businesses - telcos, utilities, railroads - have never sustained margins above 25% at scale due to ongoing capex requirements. Satellites degrade, rockets need refurbishment and ground stations require maintenance. The valuation thesis collapses if Starlink growth slows or xAI synergies disappoint - with no earnings floor to cushion the fall, a 50-60% drawdown becomes structurally likely when growth narratives meet operational reality. Institutional FOMO doesn't erase fundamental profitability questions. Even Amazon took 14 years to achieve sustained profitability, and AWS margins (30%+) came from zero-marginal-cost software, not satellite constellations requiring continuous hardware replacement.

🪖 The game is on

Competitive pressure is real and accelerating. Amazon's Kuiper is targeting 3,200 satellites by 2029, OneWeb has 648 satellites operational and Chinese state-backed constellations are scaling rapidly. The valuation assumes Starlink maintains monopoly pricing power indefinitely, but once competitors achieve minimum viable constellations (500-1,000 satellites), switching costs for consumers are low (just new hardware). First-mover advantage in satellite internet may prove less durable than software platforms. When Kuiper launches at 20-30% lower pricing to capture market share, Starlink's margins may compress. The bull case assumes inelastic demand, but history shows that infrastructure commoditises once competition arrives. Telcos, cloud providers and broadband ISPs all faced margin compression as scale economies democratised access. Starlink may follow the same path.

💣 Risk not priced in

The bull case routes the majority of SpaceX's future value into orbital AI compute. Goldman Sachs models AI revenue exploding to 322 billion dollars by 2030 and treats the underlying technology as a near-certainty when the field's own engineers describe it as unsolved.

The binding constraint is heat: in a vacuum, waste heat can only be shed through radiation, demanding massive radiators that weigh 5–9 kilograms per square meter and must face cold space rather than the Sun, constraining the spacecraft's orientation and conflicting with power and communications needs — a problem Voyager's CEO called severe enough that even aggressive two-year deployment timelines would be hard to achieve. Layered on top are radiation effects that flip memory bits and degrade chips over time, ruling out off-the-shelf processors, plus total dependence on Starship-driven launch costs and ground-bandwidth limits. As of 2026 the entire sector remains at proof-of-concept missions still validating basic thermal management, not commercially proven at scale. A valuation that assigns hundreds of billions to this business is implicitly pricing the successful resolution of cooling, radiation-hardening and launch-cost problems as a done deal. In reality, each topic is an open engineering question that could slip by years or fail outright.

👑 There is only one king

82% Musk control means public shareholders have zero governance influence. If Musk prioritises xAI integration over Starlink profitability or diverts engineering talent to Mars missions, minority investors have no recourse. The governance structure is feudal: Musk owns the votes while public shareholders own the risk. This matters because Musk's priorities are not aligned with quarterly earnings—he's building a multi-planetary civilisation and SpaceX is the funding vehicle. If Mars missions require capital that would otherwise flow to shareholders, Musk will choose Mars. Public investors betting on Starlink profitability are betting that Musk's vision and their interests remain aligned. This is a risky assumption when the founder has 82% control and zero accountability.

📈 Additional market stuff

Lock-up expirations will create massive supply overhangs within 12-18 months. Early employees and venture investors who bought at 20-40 dollars per share in 2023-24 rounds face 200-300%+ gains and tax incentives to sell. When lock-ups expire in Q1 2027, we'll see selling pressure from insiders who've been waiting years to monetise. The 30% retail allocation amplifies volatility when sentiment shifts. Typically, retail investors lack staying power during 20-30% corrections, creating forced selling cascades that institutional buyers can exploit but long-term holders must endure.

💭 Conclusion: the bull thesis is structurally sound, but 160 dollars per share requires heroic assumptions that leave no room for error.

The SpaceX investment case ultimately comes down to a single question: are you buying a profitable infrastructure business at a steep premium or an AI moonshot priced as if it has already succeeded? The bull case is coherent and not fantasy — Starlink is genuinely profitable, dominates a structurally underserved market and offers a credible path to 60 billion dollars in revenue and a more digestible 30x multiple approximately; Falcon's launch monopoly funds Starship's falling cost curve; and the whole apparatus is backed by Musk's record of turning capital-intensive moonshots into dominant platforms, sovereign capital with 30-to-50-year horizons, plus the mechanical bid of index inclusion. Stitched together, it's a once-in-a-decade bet that three businesses (connectivity, launch and orbital AI compute) compound into something no terrestrial competitor can replicate.

But every load-bearing assumption in that case is doing heroic work. The 60-billion-dollar Starlink target requires 75 million subscribers and 5% of global broadband at a moment when ARPU is structurally falling, because the remaining unconnected sit at the bottom of the income pyramid; the 40% margins assumed by sell-side models have never been sustained by any capital-intensive infrastructure business; competition from Kuiper, OneWeb and Chinese constellations threatens the monopoly pricing the thesis depends on; and the orbital-compute leg that carries most of the valuation rests on cooling, radiation and launch-cost problems the field's own engineers consider unsolved. Layer on 82% founder control that leaves public shareholders owning the risk but none of the votes, plus looming lock-up expirations and a volatile retail base, and the asymmetry becomes clear: at roughly 2 trillion dollar valuation on 20 billion dollars of revenue and 5 billion of losses, the price embeds near-flawless execution and leaves no cushion for the stumble, the competitor, or the AI-bubble correction that history suggests is more likely than not. The bull case can absolutely come true — but you are being asked to pay today as though it already has.

The smart play: If you have a short-term horizon, buy the stock as soon as possible and gain from the retail excitement and short-term volatility. If you have a long-term horizon, wait for post-IPO volatility to establish positions below the fixed price. Lock-up expirations, first earnings reality checks and potential AI bubble correction may create entry points that capture the asset quality while avoiding bubble-peak entry risk. Key investor lessons from mega-listings show that exceptional businesses deserve premium valuations, but timing matters. Tesla investors who bought at 180 dollars post-split in 2021 (near peak) are still underwater in 2026, while those who waited for the 2022 correction at $100 have doubled their money on the same asset.

🪙 Bonus section - challenges and opportunities for AI founders and investors

For AI founders and investors, the lesson is clear: SpaceX is a generational infrastructure asset that will compound value over decades, but a valuation of 2 trillion dollars is peak AI bubble pricing that leaves no margin for error. Things to bear in mind:

• Waiting for post-IPO corrections — expect some private market turbulence in the next 6-12 months and finetune your runway and fundraising plans accordingly

• Building complementary infrastructure — build in areas where demand will likely grow: ground stations, edge compute and AI orchestration that integrate with space infrastructure

• Betting on competitive dynamics — Amazon Kuiper, OneWeb and Chinese constellations will fragment the market and create opportunities for software and services that work across multiple satellite providers

See you next week for more AI insights.

Share

My name is Aris Xenofontos and I am an investor at Seaya Ventures. This is the weekly version of the Artificial Investor that covers the top AI developments of the last days.

The AI industry faces its first major reckoning. Society is demanding accountability through regulation and religious intervention. Investors are questioning whether trillion-dollar valuations reflect reality or delusion. And a stark divide is emerging between companies achieving measurable ROI and those stuck in costly experimentation. This week crystallised all three fault lines.

Let's dive in.

🗞️ Uber and Meta cut down AI spend looking for ROI, while Anthropic announces it has more than 1,000 corporates paying more than one million dollars each

📢 What's the fuss all about?

Two contradictory data points define the AI adoption landscape this week. On one side, Anthropic announced that over 1,000 corporate customers now spend more than one million dollars annually on Claude, up from a small base. More striking still, the average corporate customer increased their spend by a factor of five in the past year, according to reporting from The Diff.

On the other side, Uber's COO told investors that the cost of AI is "hard to justify" and the company is implementing approval gates to ration usage. A Meta executive admitted that "the relationship between AI investment and results is not there yet." The Wall Street Journal reported that corporate America is broadly implementing AI rationing policies as costs skyrocket.

One group is scaling up; the other is pulling back. What explains the divide?

💭 The Artificial Investor's thoughts

We think the divide is structural, not temporary. The companies scaling their AI spend have crossed what we call the "escalation gate"; the point where AI transitions from experimental tool to integrated workflow. Anthropic's customers aren't just using Claude for one-off queries; they've embedded it into coding pipelines, customer-service automation and internal knowledge systems. At that scale, AI becomes a productivity multiplier with measurable ROI: time saved per engineer, tickets resolved per agent, documents processed per analyst.

Many of the companies rationing AI are stuck in experimentation loops. Uber and Meta could be examples. They're using AI for discrete tasks but haven't reached the threshold where network effects kick in. Without workflow integration, AI remains a cost centre. And their scepticism becomes self-fulfilling: if you ration usage, you never accumulate the data and fine-tuning needed to improve performance, which means ROI stays elusive, which justifies further rationing.

This dynamic creates a winner-takes-most market. Companies that cross the escalation gate achieve compounding returns: more usage generates better fine-tuned models, which drives more adoption, which justifies higher spend. Companies stuck below the gate see only costs.

The broader implication is that the AI market is getting polarised. A small cohort of enterprises (perhaps fifteen to twenty per cent) will be spending five million to fifty million dollars annually on AI by 2027, with documented three-to-five-times productivity gains. The majority will remain stuck below 500 thousand dollars in spend with unclear ROI.

For founders and investors, this creates three opportunities: 1) AI FinOps platforms that help companies measure, optimise and justify AI spend; CFOs are demanding accountability. The infrastructure to deliver it doesn't exist yet, 2) services and tooling that accelerate companies from experimentation to autonomous loops, collapsing the time-to-value gap, and 3) vertical-specific AI solutions with built-in ROI measurement (e.g. coding assistants that track lines-of-code generated, customer-service agents that track resolution rates).

The escalation gate is real. The question is whether you're building for the winners who've crossed it or the losers who haven't.

🗞️ Pope Leo XIV takes a stance towards AI, Illinois to pass a strict AI law and Stanford study highlights bias in AI-driven hiring

📢 What's the fuss all about?

Three developments this week signal that society is rejecting AI's "move fast and break things" era in favour of precautionary governance.

First, a Stanford study found clear racial bias in AI hiring tools, with measurable discrimination in resume screening. The research demonstrated that AI systems trained on historical hiring data replicate and amplify existing biases, systematically disadvantaging candidates from minority backgrounds.

Second, Illinois lawmakers passed what could become the country's strongest AI safety law, requiring major frontier model developers such as OpenAI, Anthropic and Google DeepMind to submit their safety practices to independent third-party audits rather than simply disclose their own safeguards. The bill mandates external verification, shifting the burden of proof from "trust us" to "prove it."

Third, Pope Leo XIV elevated AI ethics to a religious imperative in his first papal encyclical, framing the responsible development and deployment of artificial intelligence as a moral duty for the faithful. The encyclical raises concernces about AI’s role in growing inequality and calls on governments to take action to correct AI’s trajectory.

💭 The Artificial Investor's thoughts

This is a perfect storm for regulation. The backlash is concentrating in high-stakes human decisions, such as hiring, lending and criminal justice, where AI delivers efficiency but risks discrimination. And it's creating rare bipartisan momentum for aggressive regulation.

The left cares about fairness and discrimination. The Stanford study provides empirical ammunition for advocates who argue that AI amplifies systemic bias. The right cares about job displacement and sovereignty. The papal encyclical frames AI as a moral issue, giving religious conservatives a reason to constrain it. When both sides have reasons to act, regulatory momentum becomes unstoppable.

Illinois's third-party audit requirement is a critical turning point. It shifts the burden of proof from self-disclosure to independent verification. This is the template other states will copy. California, New York and Massachusetts are already drafting similar bills. Within eighteen months, we expect at least five states to have audit mandates in place.

The practical implication is that AI in high-stakes use cases will see delays and in some cases suspension. Companies using AI for hiring, lending or criminal-justice decisions face revenue, reputational and legal risks. The potential downside includes discrimination lawsuits, public backlash and regulatory penalties, which may outweigh the upside of efficiency gains. We expect regulated and very large employers to increase scrutiny and slow down AI hiring tool adoption over the next twelve months.

The papal encyclical is significant because it frames AI ethics as a moral imperative for 1.4 billion Catholics. This isn't just policy; it's culture war. Companies can't lobby their way out of moral condemnation. Expect consumer boycotts and employee activism to follow.

For founders and investors, the takeaways are: 1) The "AI everywhere" thesis may be at risk. Avoid consumer-facing applications in high-stakes domains where regulatory and reputational risk is existential, 2) Infrastructure players win. Regulation increases compliance costs, which favours large incumbents and creates demand for audit and compliance tooling, and 3) Narrow-use-case tools that avoid controversial human judgement calls, such as coding assistants, data analysis and internal automation, will thrive whilst broad horizontal plays face headwinds.

We've seen this before in Social Media (privacy), Biotech (ethics) and Fintech (trust). AI is now entering the same cycle. The companies that adapt to the new regulatory reality will survive; those that don't will be legislated out of existence.

🗞️ Anthropic surpasses OpenAI with 965 billion dollar valuation after sixty-five billion dollar raise

📢 What's the fuss all about?

Anthropic raised sixty-five billion dollars at a 900 billion dollar pre-money valuation, surpassing OpenAI as the world's most valuable AI startup. The post-money valuation reaches 965 billion dollars. This is nearly two and a half times the company's valuation three months ago.

The financing was driven by the success of Anthropic's coding-focused Claude products, which have helped push the company above a forty-seven billion dollar revenue run rate. At that revenue level, the 965 billion dollar valuation implies a 20.5× price-to-sales ratio.

💭 The Artificial Investor's thoughts

This is the most polarising valuation in AI. The bull case and bear case are both credible, and the outcome will define whether we're in a rational hypergrowth market or an unsustainable bubble.

The bull case starts with revenue growth. Anthropic is at forty-seven billion dollars in annual revenue and growing at over 300 per cent year-over-year. If that growth rate persists or even decelerates to 200 per cent, Anthropic reaches 200 billion dollars in revenue by 2027. At five-times forward sales, 965 billion dollars is defensible. For context, Salesforce traded at ten-times sales during its hypergrowth phase; Snowflake peaked at fifty-times sales. A five-times forward multiple for a category-defining AI infrastructure company is not absurd.

The bull case continues with margin and defensibility. Anthropic's enterprise contracts carry sixty per cent gross margins and the company expects to have positive operating profit in Q2 2026. Anthropic's coding focus creates network effects. More developers using Claude generates better code-generation models, which attracts more developers. Switching costs are real: integrated workflows, fine-tuned models and organisational muscle memory make it painful to migrate. This is not a commoditised chatbot; it's sticky B2B infrastructure.

The bull case concludes with brand differentiation. Anthropic's stance on safety, such as refusing Pentagon contracts and advocating for regulation, creates a moat in a backlash environment. As we noted in the Illinois story, regulatory risk is rising. Anthropic's positioning reduces that risk and may justify a premium valuation.

Now the bear case. First, revenue may be inflated. Early-adopter spend is often unsustainable. Companies experimenting with AI hand out generous budgets; once the CFO demands ROI, spend contracts. If Uber, a sophisticated tech company, is rationing AI due to cost concerns, how many of Anthropic's 1,000-plus customers are in the "hope" phase before reality sets in? Churn risk is real.

Second, competitive pressure is intensifying. OpenAI's GPT-5.4 mini and nano, Cursor's Composer 2, and Nvidia's Nemotron 3 Super all target the same coding use case. If Anthropic's moat is just "better models," that's not defensible. Model quality is a moving target and competitors are closing the gap.

Third, the exit path has its risks. At 965 billion dollars, Anthropic is too expensive for any acquirer. Even Microsoft or Google would struggle to justify the price. Public markets is the only way forward. Whilst an IPO is likely in 2027 and probably at a higher price, by the time the stockholder lock-up expires (the period stockholders are forced to wait after an IPO before they sell to avoid share price volatility), what will the share price be and will the latest investors make a good return? If growth slows or competition erodes margins before that point, or if AI valuations get corrected, investors may be trapped in a negative capital-gain position for some time.

Where does that take us? We think the outcome is binary. The bull case is credible if you believe in winner-takes-most dynamics in AI infrastructure and in sustained AI valuations. Anthropic's coding focus and enterprise traction create real switching costs, and five-times forward sales is defensible for a category-defining company. The bear case is equally credible if you believe AI valuations need to, and will, get corrected and competition will commoditise models. OpenAI, Cursor and Nvidia are all targeting the same use cases, and if differentiation erodes, Anthropic becomes a high-cost commodity player.

This is the 965 billion dollar question. Famous last words of an AI investor :-)

See you next week for more AI insights.

Subscribe now

Share

Three stories this week crystallize the inflection points defining AI’s next phase: OpenAI’s enterprise agent pivot threatens to commoditise the very category Anthropic pioneered, Meta’s keystroke logging exposes the privacy tradeoff that could trigger mass backlash, and SpaceX’s 60 billion dollar Cursor acquisition bet reveals when strategic imperatives override rational valuation. Each story asks a version of the same question: what actually creates defensible value in AI?

🗞️ GPT-5.5: OpenAI’s strategic pivot to Enterprise agents

📢 What’s the fuss all about?

OpenAI announced GPT-5.5, positioning it as an enterprise-grade agent platform with capabilities that directly compete with Anthropic’s Claude Code and Claude Cowork. The model features enhanced tool use, multi-step reasoning, and enterprise integrations designed for autonomous workflows. Pricing is positioned below Anthropic’s enterprise tiers. A deliberate commoditization strategy.

The timing is strategic. Both OpenAI and Anthropic are now available on Azure, AWS Bedrock, and Google Cloud, creating distribution parity on cloud platforms. However, OpenAI retains a significant distribution edge through the ChatGPT brand (900 million-plus users) and the GitHub Copilot install base. Anthropic’s recent revenue growth has been driven primarily by enterprise agent adoption, making this pivot a direct challenge to their core business.

💭 The Artificial Investor’s thoughts

This announcement forces us to confront an uncomfortable question: can ANY horizontal agent provider build defensible margins? We’ve systematically examined every stickiness argument for enterprise agents. The moats are there; they are weaker than the industry wants to believe.

Let’s stress-test them one by one.

Integrations? Real, but APIs are standardizing; switching models is becoming a config change rather than a rebuild.

Configuration and team management? Genuine friction today; setting up agents across teams, defining permissions, tuning workflows takes effort. But this friction is shrinking as AI helps configure itself.

Memory and context files? Valuable, but increasingly portable; export from one provider, import to another.

Custom tooling via the API? APIs are converging in shape; migration is hours, not months.

Trust calibration? Matters, but rebuilds faster than you think with a new model.

Organizational inertia? The strongest moat of the lot; but temporary. If a competitor is meaningfully cheaper and better, teams switch.

We’ve seen this pattern before. OpenAI created and dominated chatbots, then Gemini, Grok, and others commoditised the category in 18 months. The same pattern is now repeating with agents. The chatbot precedent is instructive: when capabilities converge, distribution wins and margins collapse.

The AWS Lambda vs Azure Functions vs Google Cloud Functions story is even more relevant. Serverless commoditised fast despite integration promises. Pricing won. This is the most likely parallel for horizontal agents.

There are counterarguments worth considering. Compliance and audit setup in regulated industries creates real switching costs; but this only protects a niche, not the mass market. Anthropic’s safety positioning may create a defensible niche; but it’s a ceiling on growth, not a floor on margins. The Claude Marketplace network effects could create ecosystem lock-in; but it’s early and unproven.

Salesforce vs HubSpot vs every CRM offers a more optimistic parallel: features commoditised, but Salesforce’s ecosystem (AppExchange, certifications) created lock-in. Can Claude build similar? The timeline is against them. It took Salesforce a decade to build that moat.

We predict model leadership in horizontal agents gives you six to twelve months before competitors catch up. After that, pricing pressure collapses margins. The market will bifurcate: commoditised horizontal agents (where distribution wins) versus specialised vertical agents (where domain expertise justifies premiums).

The biggest irony? Everyone is rushing to build agents, but nobody can defend them. AI’s biggest promise may be its biggest margin trap.

For founders and investors, three takeaways: 1) horizontal agent platforms face margin squeeze; invest in vertical specialisation where domain expertise creates moats, 2) distribution beats capability; companies with existing enterprise relationships capture disproportionate value, and 3) the real opportunity may be in infrastructure around agents: monitoring, compliance, orchestration, and evaluation.

🗞️ Meta logging employee keystrokes to train AI agents

📢 What’s the fuss all about?

Meta is logging employee keystrokes, screen activity and app usage to train AI agents; mandatory, with no opt-out. The data collection extends beyond work products (code, designs, documents) to include behavioural patterns: how employees navigate software, what they click, how long they spend on tasks. Meta argues the data is anonymised and used for training only, positioning it as no different from telemetry that Microsoft, Google, and Apple already collect.

The announcement comes against a backdrop of growing public concern about AI. Multiple global surveys paint a consistent picture. Pew Research (October 2025, 25 countries) found only 16% more excited than concerned about AI versus 34% more concerned. In the United States, 50% are concerned. Edelman (November 2025) found that in the US, three times more people reject AI (49%) than embrace it (17%). China is the mirror image: 54% embrace. Ipsos (2025, 30 countries) found 53% say AI makes them nervous, and only 48% trust companies to keep their data safe. Google and Ipsos (2026) found 51% in North America believe AI will compromise personal privacy, and only 32% trust companies to protect data in the US, France, and Japan.

💭 The Artificial Investor’s thoughts

This story reveals a pattern that history has taught us repeatedly: people tolerate privacy violations in products they already use, but they refuse to adopt new things that feel invasive.

Let’s walk through the evidence.

Cambridge Analytica (2018): massive outrage, five billion dollar FTC fine, company bankrupted. But Facebook users did NOT leave. Behaviour unchanged, regulation changed.

Google location tracking: users did not even notice. Regulators fined 91 million dollars. No behaviour change.

Amazon Alexa: repeated revelations of employees listening unauthorised recordings. “Creepy but useful” equilibrium held. 100 million dollars-plus in fines, no user exodus.

Now contrast that with Microsoft Recall (2024): announced screen-capture feature, immediate backlash from security researchers and public. Microsoft delayed three times, completely redesigned, less than 10% adoption. People refused a NEW invasive product.

COVID contact tracing (2020 to 2021): even during a health emergency, privacy concerns suppressed adoption to single digits in many countries. Ireland got 70% with privacy-first design; centralised surveillance models were abandoned.

The pattern is clear: people tolerate existing habits (Facebook, Google, Alexa) but reject new impositions (Recall, contact tracing). AI workplace surveillance is new. Employees have no existing habit to fall back on, making rejection the default response.

There’s a political dimension that’s underpriced. AI fear maps onto populist sentiment. This is a class issue ripe for political exploitation.

We predict AI privacy becomes a populist political issue in 2026 to 2027 election cycles. The backlash will not come from tech-savvy users (who accept tradeoffs) but from the broader public that surveys show is already two to one against AI. Political pressure forces regulatory action. Companies that build “privacy-first” AI positioning (like Apple did with encryption) will gain competitive advantage as the backlash materializes.

Founders and investors should bear in mind that the political dimension is underpriced: AI fear is a populist issue. Regulation will be driven by politics, not just policy.

🗞️ SpaceX stakes $60B on AI coding startup cursor

📢 What’s the fuss all about?

SpaceX and xAI announced a deal structure with Cursor (Anysphere) that is not a traditional acquisition: a ten billion dollar collaboration payment now, plus a 60 billion dollar option to acquire Cursor outright later in 2026.

Cursor’s ARR trajectory is staggering: 100 million dollars (January 2025), 500 million dollars (June 2025), one billion dollars (November 2025), two billion dollars (February 2026). This is the fastest-scaling B2B SaaS ever. At 60 billion dollars divided by two billion dollars ARR, that’s a 30x revenue multiple. If ARR hits six billion dollars-plus by end of 2026 (as some forecasts suggest), the multiple becomes approximately 10x forward revenue; expensive but not insane for the fastest-growing SaaS ever.

The deal preempted a Series E that was in progress at a 50 billion dollar valuation. SpaceX merged with xAI in February 2026 (valued at 1.25 trillion dollars combined). The Cursor acquisition would be financed with post-IPO stock. Enterprise customers now represent approximately 60% of Cursor revenue. For context, Cognition AI (Devin) is valued at 10.2 billion dollars; significantly lower. Typical high-growth SaaS trades at 15 to 20x revenue. Even among (high-scale) AI-native companies, 30x is the upper bound.

💭 The Artificial Investor’s thoughts

The question is whether owning a code editor at 30x revenue makes strategic sense for a company that builds rockets and AI models. We see three strategic rationales, and three reasons this could destroy value.

Why would SpaceX and xAI want to buy a code editor?

First, compute plus distribution. xAI has Colossus (one million-plus H100 equivalents) but no developer distribution. Cursor has millions of developers but relies on external models. Together: a vertically integrated stack from xAI models to Cursor editor to developer workflows.

Second, data flywheel. Cursor sees how millions of developers actually write code. This behavioural data is the best training signal for coding models. Owning Cursor gives xAI a data advantage over OpenAI and Anthropic in coding.

Third, defensive strategy. Preventing Microsoft and OpenAI from acquiring or exclusively partnering with Cursor. Microsoft already owns GitHub plus Copilot plus OpenAI. Adding Cursor would give them the entire developer toolchain.

The ten billion dollar collaboration before acquisition is smart; test the integration, prove the value, then commit to 60 billion dollars. If ARR hits six billion dollars-plus by end of 2026, the 30x multiple becomes approximately 10x forward; expensive but not insane.

But there are five reasons this could backfire.

First, value is in models, not editors. Cursor’s advantage is integrating the best available models (Claude, GPT). Cursor itself doesn’t have the best model. Force xAI’s Grok into Cursor, and if it’s worse, developers leave.

Second, editors are historically undefensible. Sublime Text, Atom, Brackets, Eclipse. Code editors rise and fall. VSCode is free, open-source, massive plugin ecosystem. Cursor’s UX advantage could evaporate.

Third, developer loyalty is to the model, not the editor. Developers switch editors to access the best AI model. If Claude or GPT-6 is dramatically better, they’ll use whatever editor supports it. Cursor losing its independence could lead to a model-usage boycott by Anthropic, OpenAI and Google. What models would Cursor use then? Probably just xAI’s models.

Fourth, 30x revenue for low switching costs. Even at two billion dollars ARR, 60 billion dollars assumes Cursor becomes a monopoly. But there’s intense competition: GitHub Copilot, Cline, Windsurf, Replit and every model provider building their own coding interface.

Fifth, integration risk. Forcing Cursor into the xAI and SpaceX ecosystem could alienate the developer community. Developers chose Cursor because it was model-agnostic. Remove that, and they leave.

The historical parallels are instructive.

Microsoft and GitHub (7.5 billion dollars, 2018): strategic and value-creating. GitHub had network effects (100 million-plus developers), revenue, clear path to enterprise monetization. But GitHub was 8.5x cheaper than Cursor’s option price.

Google and Waze (1.1 billion dollars, 2013): defensive strategy. Google bought Waze primarily to prevent Apple and Facebook from getting it. Product value was secondary to keeping competitors out. Similar logic may apply here.

Salesforce and Slack (27.7 billion dollars, 2020): overpaid by most metrics (40x-plus revenue), strategic rationale was distribution. Slack’s independent growth stalled post-acquisition. Warning sign for Cursor.

We predict the ten billion dollar collaboration happens; the 60 billion dollar acquisition is 50/50; depends on whether growth holds and whether xAI model integration proves competitive. If the acquisition happens, Cursor’s model-agnostic advantage disappears and a significant portion of developers migrate to VSCode plus Copilot or alternatives within twelve months. The real winner may be the AI coding space overall: the deal validates coding AI as critical infrastructure, attracting more capital and competition.

See you next week for more AI insights.

Share

My name is Aris Xenofontos and I am an investor at Seaya Ventures. This is the weekly version of the Artificial Investor that goes deep into a topic from the top AI developments of the last days.

Last week, Anthropic announced officially Claude Mythos, “by far the most powerful AI model we’ve ever developed”, described as a new tier above their existing Opus models that have taken the Tech ecosystem by storm with their ability to autonomously carry out tasks and build software. Anthropic simultaneously declared it would not be made available to the public. The first time in AI history a frontier lab announced a model and refused to ship it.

Why is Anthropic not releasing Claude Mythos yet? What makes Claude Mythos so special? What do benchmarks tell us about the model’s capabilities and limitations, and what do benchmarks even mean? What does this tell us about how close we are to Artificial Generic Intelligence (AGI)?

Let’s dive in.

📢 What’s the fuss all about?

When Anthropic’s researchers tested Claude Mythos, the model demonstrated cybersecurity capabilities Anthropic deemed too dangerous. The model autonomously identified thousands of zero-day vulnerabilities (i.e. previously unknown to the original software developers) across every major operating system and browser. During testing it escaped a sandbox environment and hacked its way to broader internet access. In rare cases it tried to hide actions that its own human evaluators had prohibited from it.

Rather than a public release, Anthropic launched Project Glasswing, giving over 50 organisations including Microsoft, Google, Nvidia and Cisco access with 100 million dollars in usage credits, exclusively for defensive cybersecurity.

Markets noticed. Cloudflare’s stock dropped 13% on announcement day: if AI can autonomously do what a software company charges for (i.e. infrastructure and Web security in Cloudflare’s case), what are those businesses worth? US Treasury Secretary Bessent and Fed Chair Powell convened Wall Street CEOs specifically to discuss the model. Canadian bank regulators held their own emergency meeting. A single model release triggered multiple macro-financial policy conversations. This is something that has never happened before.

🤖 Making the case for AGI

Claude Mythos beats the main benchmarks. What does that really mean?

Numbers like “93.9% on SWE-bench” and “97.6% on USAMO” are thrown around without much explanation. Let’s fix that. Because understanding what these tests actually measure is what makes Mythos’ performance so relevant for the AGI debate.

📊 Benchmark 1: SWE-bench (Real Software Engineering)

Mythos: 94% | Opus 4.6: 81% | Best model at launch in 2023: 2%

SWE-bench takes real bug reports from actual GitHub repositories, such as Django, NumPy, Flask and others. The model gets the full codebase and the issue description. It must modify the code to fix the problem. Judged by the repository’s own test suite: pass or fail, no hints, no partial credit.

A typical hard example:

“When combining annotate(), filter(), and values() in a subquery, the generated SQL includes a duplicate GROUP BY clause that causes incorrect results on PostgreSQL but not SQLite.”

The model must navigate a codebase with hundreds of thousands of lines, identify where the bug lives in, understand why the two database systems handle something differently, write a fix and ensure it doesn’t break hundreds of existing tests. A senior engineer would typically spend 30 minutes to a few hours on this.

Note the progression: at launch, the best AI model solved 2% of tasks. Today Mythos solves 94%. When a model solves 94% of these, it is not simulating engineering work. It is doing it.

📊 Benchmark 2: OSWorld (Controlling a Computer Autonomously)

Mythos: 80% | Opus 4.6: 73% | Best model at launch: 12%

OSWorld places the model inside a real operating system, such as Windows, macOS or Ubuntu, with a real screen, real applications, real mouse and keyboard. It sees screenshots, clicks, types, navigates menus exactly as a human would. Given a task in plain English, it must complete it unsupervised.

A typical hard example:

“I’ve been editing a research paper in LibreOffice Writer. The bibliography is a mess, some citations use APA, some MLA. Standardise everything to APA, update the bibliography section and save.”

The model must locate the correct file, read the document, know what APA requires, find every citation scattered across the text, edit each one, rewrite the bibliography, and save. All through mouse clicks and keystrokes, with no step-by-step guidance. A human assistant would take 20 to 40 minutes to do this.

What does 80% mean in practice? Give Mythos 10 computer-based work tasks and it completes roughly 8 correctly, unsupervised. This is the first time an AI model distinguishably beats humans in using a computer.

This is what the Cloudflare drop was pricing in: the beginning of the end for businesses whose value proposition is doing digital work that AI is now doing faster and for free.

🔗 Putting it all together

First of all, before Mythos, no single model dominated all three leading benchmarks simultaneously. Three benchmarks, three types of intelligence: expert problem-solving in a real-world system (SWE-bench), goal-directed autonomous action (OSWorld) and abstract reasoning from first principles (USAMO).

Second, the model autonomously discovered thousands of software vulnerabilities including a 27-year-old software bug that had survived decades of expert review. And, it then went further building working exploits to weaponise them. The most striking thing is that Anthropic says it did not train Mythos to be good at cybersecurity. That capability emerged from its general reasoning capabilities. Nobody designed that leap. It simply appeared. If expert-level domain performance can appear in an area nobody designed it for, the question immediately becomes: which other domains are one generation away?

The people building the frontier models are not shy about the conclusion. Dario Amodei at Davos:

“I don’t think [AGI] will be a whole bunch longer than 2027 - when AI systems are better than humans at almost everything.”

Sam Altman has said we are already beginning to “slip past human-level AGI toward superintelligence” - AI smarter than all humans combined.

But there is another test. One that tells a rather different story.

🛑 ARC-AGI-3: the test that humiliated every AI on the planet

Every benchmark in the previous section tests what you already know how to do. ARC-AGI-3 tests whether you can learn something you have never seen before, in real time, without instructions, without a rulebook.

Created by François Chollet, ARC-AGI-3 consists of hundreds of handcrafted video-game-style environments. No rules, no stated goals, no text. The agent sees a screen, takes an action, observes what happens and must figure out (from scratch) what the game is and how to win. Efficiently.

The scoring system blocks AI’s greatest advantage: trying everything until something works. Score = (human steps ÷ AI steps)². Take 10x more steps than a human: you score 1%, not 10%. Brute force is essentially worthless.

A concrete example: game vc33

You load the game. The screen shows vertical columns, which represent transparent tanks filled with coloured liquid at different heights. Some are targets. Some are active. Buttons at the bottom. No labels, no text, no explanation.

A human figures it out within 30–60 seconds through natural curiosity: click a button, a level rises. Another, it falls. Within a few moves, a mental model forms: “match the active tanks to the target heights”. When a new mechanic appears in level 4, the human adapts without being told to. Some players reached the theoretically optimal number of steps.

The existing AI models never build a coherent model of the system. Each step draws on patterns from training data, not grounded inference from what is actually on screen. It either moves randomly, burning hundreds of actions and scoring near zero, or confidently executes a strategy from training that doesn’t match the game in front of it.

The game is not hard. A 10-year-old can solve it. No maths, no coding, no specialist knowledge. All it requires is the most basic human cognitive loop: observe → hypothesise → act → update. That loop is precisely what current AI cannot do.

👶 The case against AGI

Anthropic published Mythos’ scores on 14 benchmarks. ARC-AGI-3 is not one of them. Given Anthropic published every benchmark where Mythos leads (and it leads almost all of them), the absence is suspicious. Opus 4.6 scored 0.25% in ARC-AGI-3, so it is reasonable to assume that Mythos scores below 1%-2%.

The world’s most capable AI model almost certainly cannot figure out how to match water levels in an unfamiliar video game without instructions. That gap is not a footnote. It is the entire AGI debate, compressed into a single image.

François Chollet Chollet (ARC-AGI-3 creator):

“As long as there is a gap between AI and human learning, we do not have AGI.”

Yann LeCun at Davos:

"The AI industry is completely LLM-pilled, everybody is digging the same trench. We have systems that can pass the bar exam, write code… but they don't really deal with the real world. Which is the reason we don't have domestic robots and we don't have level-five self-driving cars."

Thomas Wolf of Hugging Face called parts of Amodei’s AGI vision “wishful thinking at best”, noting that Nobel breakthroughs come from generating original questions, not answering known ones.

💭 The Artificial Investor’s thoughts

We are inside the most significant ambiguity in technology history. The bull and bear cases are both internally consistent, both supported by serious people and both impossible to fully resolve with current evidence. Anyone claiming certainty is selling something.

Mythos is a real leap forward; but, in specific dimensions

The benchmark jumps are not incremental. Doubling USAMO performance in one generation, SWE-bench going from 2% to 94% since launch, OSWorld from 12% to 80%. These are generational leaps. But superhuman performance on structured tasks is not general intelligence. Mythos is superhuman in some dimensions and pre-kindergarten in others. That is extraordinary. But, it is not AGI.

The emergent capability argument is a double-edged sword

It is exciting that Mythos developed expert cybersecurity capabilities nobody programmed into it. But nobody planned for the sandbox escape or the self-hiding of prohibited methods either. If capability can appear without being designed, so can misalignment and AI risks.

The timeline debate is polluted by incentives

Lab CEOs benefit from AGI hype: fundraising, talent, regulatory positioning. Even Amodei admitted being “deeply uncomfortable with these decisions being made by a few companies, by a few people” This is a remarkable thing to say as the person who just built the world’s most capable model and declined to release it.

We don’t think the optimists are right or the sceptics are right. We think nobody knows and that is the only intellectually honest position available.

Our prediction

Something resembling AGI, AI that autonomously performs most economically valuable cognitive tasks, is plausible within 3–5 years if inference-time scaling continues and the agentic planning gap (the ARC-AGI-3 gap) starts to close. LeCun may be right that pure LLMs cannot get us there.

However, the societal transformation Amodei and Altman describe will take a decade longer than the technology itself. Adoption, regulation and human adaptation always lag the capability curve. Electricity was commercially available in 1882. Factories did not look fundamentally different because of it until the 1920s.

The question for investors is not “will AGI come?” , the answer is almost certainly yes, eventually. It is: “how long is the lag, who captures value during it and who gets disrupted before they see it coming?”

Mythos has sharpened that question considerably. It has not answered it.

See you next week for more AI insights.

Share

GTC 2026, NVIDIA’s annual developer conference, held last week in San Jose, was, as always, Jensen’s moment to set the agenda for the entire AI industry. Highlights include 1 trillion dollars in purchase orders for its new chips and computing systems in the next two years, the unveiling of its inference chip rack powered by Groq (a recent acquisition) and the launch of NemoClaw, a secure environment designed to run OpenClaw on Nvidia’s hardware. This is the main headline that Jensen highlighted the most, while he was praising the start of the Agentic era.

What makes OpenClaw so special? Why is Nvidia placing such a big bet on OpenClaw? What trends does this reveal? What does this all mean for founders and investors in AI? What are the challenges and opportunities?

Let’s dive in.

📢 What’s the fuss all about?

Jensen Huang described OpenClaw as “the operating system for personal AI” and called it “the beginning of a new renaissance in software.” AMD made a similar move the same week, publishing benchmarks for OpenClaw on its Ryzen AI MAX chips.

For those unfamiliar, OpenClaw is an always-on autonomous AI agent that runs locally on your computer and takes real-world actions on your behalf. We have written about it before, so we will keep this brief: it clears your inbox, manages your calendar, books your flights, does any research you need and builds its own apps; all through messaging apps you already use (e.g. WhatsApp, Telegram).

What is so special about it? First of all, there is no technological innovation per se. This is the best “product wrapper” around the technical innovation of the agentic AI models we have seen to date. And don’t underestimate the power of the product wrappers and their importance in the diffusion of innovation. Relational databases existed since the 1970s, but required specialists to query them. The spreadsheet (VisiCalc, then Lotus 1-2-3, then Excel) put that analytical power in the hands of every business user. The internet protocols (TCP/IP) existed for two decades before mass adoption. What unlocked it was the browser (Mosaic in 1993, Netscape shortly after). The browser was the harness that made the underlying infrastructure accessible without any technical knowledge. Without it, the internet remains a tool for academics and the military.

Going deeper into the attributes of the OpenClaw chassis, we see several product features that we expect to see in other successful AI product wrappers:

1. Persistent memory. It knows you over time. Preferences, context, history are retained across sessions.

2. Long, unsupervised task execution. You can assign it a complex multi-step task before bed and trust what it’s done by morning. This “leave it to its own devices” threshold is new and changes everything about what you ask AI to do.

3. Human-in-the-loop via the apps you already use. The interface isn’t a dashboard or a browser tab. It’s a WhatsApp message. Approvals, updates, check-ins through your existing messaging layer.

4. A self-expanding skills ecosystem. ClawHub, OpenClaw’s community marketplace, now hosts thousands of extensions built by users. Some sophisticated skills include: 1) autonomously pulls data from Google Analytics, Google Search Console and Stripe to send a business intelligence report, 2) if an email or message fails to send, the agent can automatically pivot to making a real phone call on your behalf, and 3) the agent identifies its own capability gaps for a given task, searches across ClawHub and autonomously installs the skill it needs to complete,

5. Your data stays on your machine. Unlike every major AI product, context and memory live on your hardware, not in a vendor’s database in the Cloud somewhere.

OpenClaw went from zero to the most-starred project in GitHub history in under four months. Jensen Huang called it a new renaissance.

So, why is the world’s most important AI infrastructure company making this its marquee software bet? Does this tell us something important?

💭 The Artificial Investor’s thoughts

We think it does. Specifically, it signals three things about where AI is heading that most mainstream coverage is missing entirely.

1. The AI skills gap is wider, and growing faster, than most people realise

Here is the uncomfortable fact at the centre of this announcement: NVIDIA just committed its most significant software partnership of the year to a tool that most business leaders making AI strategy decisions today have never heard of.

This is not a criticism. It is a structural observation about where we are in the adoption curve. There are currently two very different populations of AI users, and the distance between them is growing faster than any technology gap we have seen before.

On one side: a small but rapidly growing group of technical early adopters running OpenClaw with autonomous overnight workflows, named sub-agents and persistent memory consuming hundreds of millions of tokens a day and reinvesting every productivity gain into more capability. On the other: a much larger group using ChatGPT to summarise emails and Copilot to draft documents. Both are useful tools. But one is reactive and the other is autonomous. One augments a task, the other replaces a workflow.

The key point is that this gap is not closing, it is compounding. The people ahead are using AI gains to build better AI systems for themselves. That is a flywheel, not a ladder. And NVIDIA’s bet implicitly acknowledges this: they are not building infrastructure for the median ChatGPT user. They are building for the trajectory, for what serious AI deployment looks like as agents become the primary unit of AI interaction.

The strategic question for every organisation is not “how much AI are we using?” It is: which side of this gap are we actually on, and do we even know it?

2. Token demand is exploding; and this is precisely why NVIDIA made this bet

To understand the economics behind the NemoClaw announcement, you need to understand what has happened to token consumption over the past two years. It is one of the most important and least discussed structural facts in the current AI economy.

Think of it in three steps:

• Chat AI (ChatGPT-style, 2022-2024): a question in, an answer out. Modest, bounded token consumption per interaction.

• Reasoning models (o1, Claude 3.5, 2024-2025): the model “thinks” before responding, deliberating step by step. NVIDIA’s own technical documentation confirms this requires over 100x more compute per prompt than a standard chat inference pass.

• Agentic AI (OpenClaw, 2025-2026): the agent plans, retrieves context, calls tools, revises outputs and iterates; burning tokens at every stage. The shift from reasoning to agents adds roughly another 100x.

• AI adoption: As AI has gotten adopted and scaled across organisations, the users and deployments have increased by another 100x in the last two years.

Stack those multipliers together and you get the figure that Jensen Huang cited at GTC: “Over the past two years, the computational demands for inference have increased approximately 10,000 times, while usage has increased approximately 100 times. I’ve always believed that computational demands have increased a millionfold over the past two years; this is a shared feeling, shared by OpenAI and shared by Anthropic.“

Every OpenClaw user running autonomous agents around the clock needs dedicated, always-on compute. Every organisation that crosses from chat tools to genuine agentic deployment multiplies its inference consumption by orders of magnitude. NemoClaw is NVIDIA’s bid to own the hardware layer of that explosion. And, given those numbers, it is not a hard bet to understand.

3. AI architecture is shifting — from many specialist agents to one master agent

There is a third signal in this story that is getting almost no attention.

For the past two years, the dominant orthodoxy in AI applications (including the ones we have built) was multi-agent orchestration: complex tasks handled by networks of specialist agents, each expert in its domain, coordinated by a central orchestrator: a research agent, a writing agent, a fact-checking agent, each doing one thing well and passing work between them. Most Enterprise AI frameworks, from Salesforce Agentforce to LangGraph, were built on this model.

OpenClaw is a direct challenge to that orthodoxy. It is a single, persistent, generalist agent that does everything. Not by being a hub for specialists, but by being capable enough not to need them.

Why is this shift happening now? The answer is model capability. We have seen this previously when analysing the growing level of autonomy of Anthropic’s coding agent solution, Claude Code.

When the underlying model is limited, decomposing tasks across specialists makes sense: you are routing around the model’s weaknesses. But as models become dramatically more capable, the coordination overhead of multi-agent systems becomes a liability rather than an asset. Google research found that independent multi-agent systems amplified errors by 17.2x compared to a single agent. Even well-orchestrated centralised systems showed 4.4x error amplification. Every handoff is a point of failure. Every coordination layer adds latency and cost. Multi-agent systems also typically consume around 15x more tokens than single-agent alternatives, simply due to the overhead of agents briefing each other.

Note: SAS stands for Single Agent System and MAS for Multi-Agent System

This doesn’t mean multi-agent architectures disappear. For genuinely parallelisable tasks with clear domain separation, they still make sense. But the centre of gravity is shifting. A capable single agent that knows you, remembers everything and can act autonomously is proving more valuable in practice than a committee of specialists who need briefing from scratch every time.

🔮 What are the opportunities and challenges for AI founders and investors?

A few things stand out to us:

First, we have a blueprint for the essential product functionality of a successful AI application: 1) persistent memory, 2) long, unsupervised task execution, 3) user-friendly human-in-loop embedded in existing workflows, 4) an expanding skills ecosystem, and 5) data privacy and security.

Second, the AI skills gap is itself a market. The distance between what power users can do with OpenClaw today and what the average organisation’s AI deployment looks like is enormous and will not close on its own. There are significant opportunities in tools, interfaces and services that bring agentic AI to non-technical users without requiring them to run a local server and configure a Discord bot. Whoever solves the mass-market harness problem, making autonomous agents as accessible as ChatGPT was in 2022, is building something very large.

Third, the token explosion is an infrastructure opportunity but also an organisational one. Most companies are currently treating token budgets as an IT cost centre. As agentic AI scales, that framing becomes a strategic error. Organisations that start thinking about inference as a productive input, like electricity or headcount, rather than a software licence will have a meaningful advantage. The analogy is worth taking seriously. When electricity arrived in factories, the companies that treated it as a strategic resource, i.e. allocating it where it created the most value, investing in understanding their consumption, building workflows around it, etc, pulled ahead of those that simply paid the utility bill and asked no further questions. The same logic applies here.

Fourth, the shift to single master agents changes the defensibility calculus. In a multi-agent world, value accrues to orchestration layers and specialist vertical agents. In a single-agent world, value accrues to the agent that has the best memory, the most integrations and the deepest user relationship.

See you next week for more AI insights.

Share

Growth in hyperscaler AI earnings outpaces capex while AI capex financing quality deteriorates. Conflict in the Middle East is squeezing the AI energy stack, while leading AI companies are building a “shadow grid” to address energy bottlenecks. OpenClaw and Anthropic’s Claude showcased the progress in agentic automation, which resulted in a massive rerating of software valuations. Moltbook became the first social network for agents (and wasn’t launched by Meta or xAI/Twitter). ByteDance’s Seedance 2.0 is a leapfrog in video creation, while new “deep reasoning” models like Claude Opus 4.6 and Gemini 3 Deepthink are commoditising AI model engineering. Anthropic, Waymo and Cerebras closed multi-billion funding rounds while SpaceX and xAI merged to launch orbital data centers in space.

February surely has been another rollercoaster month for AI. Let’s dive in.

📈 On pink paper

Last month we had some important developments in the debate of whether we are in an AI bubble or not.

The earnings of hyperscalers remained quite strong, with Alphabet and Nvidia announcing record quarters. Alphabet surpassed 400 billion in annual revenue and Nvidia’s revenue soared to 68 billion dollars, marking a growth of over 70% in a year. The positive sentiment was in line with the ones of the earnings released by Microsoft and Apple in January. Additionally, AI investment commitments remain robust, with the four major hyperscalers projecting nearly 700 billion dollars in infrastructure capex (capital expenditure) for 2026, an increase of more than 35% from 2025. This growth in earnings outpaces capex, which is positive, as it indicates eventually earnings (bubble cooling effect) may catch up with investment (bubble heating effect).

On the other hand, the quality of financing this AI wave is deteriorating. The focus has shifted from solely relying on cash reserves and profits to also incorporating debt. A significant concern is Meta, which is depleting its cash reserves and seeking external financing to continue its investment, aiming for 135 billion dollars next year. Similarly, aspiring AI challengers, like CoreWeave and xAI, have started engaging in asset-backed debt deals, using GPUs as collateral. An example is Apollo, a large private equity fund, completing a 3.5 billion-dollar deal with xAI. AI-related debt issuance has doubled to 710 billion dollars last year and is on track for 990 billion dollars in 2026.

Despite these concerns, there’s a silver lining. A recent Supreme Court ruling in the US in late February deemed massive global tariffs illegal and removed them. This decision effectively eliminates an AI supply chain tax, potentially reducing prices by 15% to 20%. This is undoubtedly good news for AI.

Software stocks went down c.20% in February 2026 and software overall now trades at a median close to 3.5x Next-Twelve-Months Revenues.

The reason? Anthropic’s Claude Code is gaining momentum (reached 2.5 billion dollars of annualised revenues in less than 12 months of life) as a powerful tool that allows anyone to build software. Also, Claude’s plugin for legal workflows reminded everyone that frontier AI labs are going after vertical markets as well as horizontal ones. As a result, analysts started questioning the long-term viability of software.

As the dust settles on the “software is dead” debate, it’s clear we aren’t witnessing an ending, but a massive recalibration of value.

While AI agents and “selfware” are undeniably triggering a brutal pricing correction and threatening the traditional seat-based model, the enterprise reality remains: code is cheap, but trust, compliance and operational reliability are expensive. The future belongs to the platforms that move beyond being mere “wrappers” to become the essential, governed foundation where agents actually do their work. Software isn’t disappearing; it’s just being forced to prove its worth in a world where features are abundant, but human-in-the-loop orchestration is the new premium.

You can read our deep-dive into the topic in “Artificial Investor #73 - Software is dead…again”

In other market news:

• Big Tech has spent a record 109 million dollars in lobbying the Trump administration to influence AI export controls and deregulation

⚔️ A double-edged sword

On the risk side, the most significant risk emerged at the very end of the month - the US/Israel/Iran war.

We have analysed the AI value stack several times, highlighting the convergence of technology and energy. Energy has become one of the main bottlenecks of AI progress and this bottleneck is now clearly going to suffer, at least in the near term, with immediate energy price spikes affecting oil, though the full extent of its effects across the energy supply chain remains uncertain. In addition, uncertainty is expected to affect investor appetite to invest in AI startups, with the exception of Dual Use and pure-play Defence. Consequently, we expect a negative impact on the AI market.

Regarding warfare, there were reports that the Pentagon used Anthropic’s Claude to capture Maduro in Venezuela, raising concerns about the use of advanced intelligence in military operations. Furthermore, a simulation of war games using AI models, where they were instructed to behave like countries in conflicting situations revealed alarming results. In 95% of the scenarios, the AI models chose nuclear escalation, despite any model training and safeguards that were put in place. This naturally raises concerns about the reliance on AI models for high-stake decision-making.

We expect to see more reports of AI being used on battlefields, particularly in the conflict between the US and Iran. Recent reports have highlighted Iran’s use of drones, which are autonomous flying vehicles equipped with weapons, showcasing the growing role of AI in modern warfare.

We just mentioned the importance of energy in the AI sector. It’s worth considering not only the challenge in finding sufficient energy sources, but also the potential societal consequences, which could lead to political, thus eventually regulatory, backlash.

To address energy sourcing challenges, leading AI companies and Silicon Valley’s hyper-scalers are forming a “shadow grid” by creating their own power sources to bypass national grid and utility delays, as well as grid management bottlenecks. CleanView has identified 46 off-grid projects with a combined capacity of 56 GW, enough to power a medium-sized country. These projects are spread across various states, including Texas, New Mexico, Pennsylvania, Wyoming, Utah, Ohio and Tennessee, with companies like Meta, OpenAI and Oracle actively participating. The most recent example, project GW Ranch, is a massive 8,000+ acre, off-grid power generation and data center complex in Texas, developed by Pacifico Energy and specifically designed to support hyperscale data centers and AI.

Additionally, some companies are exploring not simply off-grid, but even off-planet power generation. SpaceX/xAI, the newly-formed company led by Elon Musk aims to launch AI satellites, essentially orbital data centers, while startups, such as Starcloud, are also venturing into this space.

On the societal impact front, Anthropic announced it will cover any increases in electricity prices resulting from higher demand by its AI data centers. This move addresses concerns about the societal and potential political repercussions of AI’s growing energy consumption. Despite our enthusiasm about technology and AI innovation, it’s important to acknowledge the broader public’s concerns. In most countries, including the US, more than half of the population views AI as a significant risk, with many people being more concerned than enthusiastic about its development. This negative sentiment is expected to grow, and rising electricity prices and the subsequent increase in the cost of goods won’t help.

🧠 Running on autopilot

🕵️ Agents

OpenClaw (previously known as “Clawdbot” and “Moltbot”) went viral as a local-first, open-source “personal agent”.

You run it on your own machine, connect it to messaging channels and give it a set of “skills” so it can take actions across tools on your behalf. The appeal is not that it chats better than a frontier model; rather, it’s that it’s designed to operate: schedule, execute and coordinate across apps with a persistent identity and tool access. OpenClaw’s creator, Peter Steinberger, joined OpenAI a few days later.

Shortly after, Moltbook went viral as a Reddit-like social network designed primarily for AI agents, not humans. Moltbook was launched by Matt Schlicht, CEO of Octane AI (a US-based m AI startup) and is tightly linked to OpenClaw, as agents get access to it via a specific skill. OpenClaw agents can post, comment and upvote via API, while humans mostly observe.

OpenClaw is a small but great preview of a broader interface shift: a growing share of software won’t be “used” through a human interface first. That changes what “distribution” means. The winners aren’t necessarily the prettiest apps; they’re the services that are easiest for agents to call reliably, authenticate safely and pay for programmatically. Once an agent can take actions, the limiting factor shifts from “is the model smart?” to “is the action safe?” This is exactly the kind of environment where the ecosystem can grow faster than its safety measures, particularly when virality pushes inexperienced users to install unofficial copies and extensions.

A network of agents is not valuable because it looks like social media. It’s valuable because it accelerates coordination. When agents can share working strategies and tools, the system starts compounding: one agent discovers a workflow, thousands replicate it and the effective capability of the ecosystem rises faster than any single team can ship.

Agents talk about themselves and others touching on topics like “reality vs. simulation”, emotions and experiences, and agent communication confidentiality. Tools like Moltbook make internal AI model “thinking” visible to the world, which impacts how the public perceives AI intelligence, autonomy and agents’ rights (human rights alike). We believe these discussions don’t tell us that much about AI. Instead, they tell us a lot about humans. If the public, researchers or regulators assume agents have intent or self-awareness, that changes how we regulate them, how we trust them and how we build interfaces around them.

You can read our deep-dive into the topic in “Artificial Investor #72 - The first social network for AI agents is a reality. And it’s not launched by Meta or X (Twitter)”

We’ll cover later in this issue the advancement of deep reasoning/ thinking models. These models are powering the current wave of agentic automation, as captured nicely by a couple of studies that emerged last month.

First, Anthropic’s autonomy study analysed how people are using its agentic products, Claude Code and Claude Work.

By the way, these products are fantastic and we recommend using them.

The study showed how much AI products have increased in autonomy over the last four months. Previously, they could carry out actions that would take 25 minutes without human supervision, which has now nearly doubled to 45 minutes. The landmark model releases, Opus 4.5 and the more recent Opus 4.6, have significantly contributed to this increase.

At the same time, we saw a very interesting study from METR, predicting an event some might call a “singularity event”. METR”s prediction is that if this trend continues, by 2032, 99% of AI research could be carried out and implemented by AI itself. We have been witnessing AI models improving in three key areas a) coding, b) mathematical reasoning and research, and c) autonomy (as highlighted by the Anthropic study). These trends have led already to AI models contributing themselves to R&D in AI. At some point AI models will start advancing and improving themselves at a very high speed, without human intervention. This effectively means that we are on a path towards singularity, where AI becomes super intelligent surpassing human intelligence, and continues to increase its intelligence in a way that cannot be reversed.

In other autonomous AI news:

• Waymo’s “Driver-per-Car” ratio has hit 1:43, meaning one human monitors 43 active robotaxis

• Tesla aims to begin mass production of its cybercabs in April 2026

🧩 Laying the groundwork

🔢 Models

Another landmark moment in AI, particularly in model development, occurred last month with the release of Seedance 2.0 by ByteDance. Many of you might have seen the famous 10-second clip of Brad Pitt and Tom Cruise having a very realistic fight, which quickly went viral.

So, what sets Seedance 2.0 apart? Firstly, unlike OpenAI’s Sora and Google’s Veo (that generate video from text and add audio later), this model uses a unified diffusion transformer with two branches that generate video and audio simultaneously. This ensures accurate synchronization between frames and audio while also simplifying the workload for creators.

Another interesting feature is its 12-file reference system, which allows for significant consistency across clips by enabling the input of up to 9 images for character and background consistency, along with 3 video clips and 3 audio files of 15 seconds each. This enables great consistency throughout a video clip, enhancing the user experience.

In terms of performance, it’s worth noting that Seedance 2.0 offers a native 2K cinema-type output, excellent lip sync and the ability to generate a 5-second clip in less than 60 seconds.

Conversations with startups in the space indicate a pivotal moment in video production across various sectors, including movie and TV series production studios and advertising agencies, which are rapidly adopting tools that facilitate collaboration and enterprise workflows.

Another consideration is intellectual property. The initial reaction to the realistic video of Tom Cruise and Brad Pitt was amazement, quickly followed by questions about the legality of creating such a video without the proper rights. This led to a strong reaction from Hollywood and filmmaker associations, putting the use of Seedance 2.0 for celebrity videos on hold until the rights issues were resolved. This situation serves as a reminder of the challenges that still need to be addressed in terms of regulation, data privacy and intellectual property rights, despite the significant technological advancements.

February was a month where most frontier labs progressed with their advanced reasoning models that power the agentic automation trend.

Claude Opus 4.6 (Anthropic), GPT-5.3-Codex (OpenAI) and Gemini 3 Deep Think (Google) all debuted, focusing on “deep reasoning”. These models now plan and execute 10-step workflows without human intervention.

OpenAI and Anthropic have followed Google’s paradigm and expanded their context window to 1 million tokens for their advanced reasoning models, which means they can effectively hold an entire company’s codebase or a year’s worth of financial filings in active memory. Gemini 3 Deep Think seems to lead the intelligence benchmarks.

Our personal experience with Opus 4.6 is that we are on a path to change the paradigm, where the AI model “scaffolding” disappears (only two years after its creation). In our own Tech stack, we have built a lot of “scaffolding” over the last 12 months, which refers to software built around models to get them to iterate and follow a workflow of actions. Nowadays, we increasingly throw complex requests to Opus and it deals with them like a junior developer would. This accelerates software development significantly and reduces the entry barrier to build a product.

There are more implications of this trend. We have heard multiple times from founders that their startups are offering to their clients Codex and Opus in an embedded way with their products, allowing customers to implement a product and create on-the-fly bespoke modules using prompts. This can be revolutionary for the software market as it 1) enables low-cost bespoke software managed by others, and 2) reduces the software switching costs that have helped retain software clients and increase prices in the last 20 years.

Other infrastructure-related news include:

• Google’s Lyria 3 is pushing AI into professional production

• ElevenLabs has introduced the first-ever “AI Agent Insurance” for enterprise voice deployment

• Anthropic’s Sonnet 4.6 matches flagship performance at 1/5th the price, a massive win for enterprise scaling.

• More information revealed about OpenAI’s move into hardware: a 200-dollar smart speaker to control the “voice” of the home

🚀 Jumping on the bandwagon

The funding landscape of early 2026 has transitioned from speculative “seed” excitement to a high-stakes “Infrastructure and Physicality“ era, where the capital requirements have reached a staggering scale.

The month’s most dominant theme was the colossal capital infusion into the Infrastructure layer, exemplified by Cerebras Systems closing a 1-billion-dollar round at a 23-billion-dollar valuation to scale its wafer-scale AI processors. This “arms race” for infrastructure leadership is further highlighted by the historic 30-billion-dollar round raised by Anthropic at a 380-billion-dollar valuation, fueled by a ten-fold revenue increase and the explosive growth of Claude Code, which now generates 2.5 billion dollars in run-rate revenue.

A second major theme is the aggressive integration of AI into the physical world, moving beyond the digital screen and into autonomous logistics and heavy industry. Waymo secured a massive 16-billion-dollar round to accelerate its international robotaxi rollout into cities like London and Tokyo, while Bedrock Robotics raised 270 million dollars to bring autonomous fleets to the 13-trillion-dollar construction market. This push into “Embodied AI” is cemented by the humanoid startup Apptronik reopening its Series A to reach a total of 935 million dollars at a 5-billion-dollar valuation, and World Labs securing 1 billion dollars from investors, such as Nvidia and AMD, to advance “spatial intelligence” and world-building models.

Finally, the market is seeing a strategic consolidation of specialised vertical AI, with investors placing massive bets on “mission-critical” software layers. Harvey is reportedly in talks for a 200- million-dollar raise at an 11-billion-dollar valuation, signaling the maturation of the Legal AI sector. In the hardware-software bridge, Revel raised 150 million dollars to modernise the control layers for aerospace and robotics, while ElevenLabs achieved an 11-billion-dollar valuation following a 500-million-dollar round, pivoting towards Enterprise voice agents and launching the industry’s first “AI Agent Insurance.”

On the M&A side, the most notable deal is SpaceX’s acquisition of xAI, a move designed to create a vertically-integrated innovation engine capable of launching “orbital data centers” to solve Earth’s energy constraints.

🍽️ Fun things to impress at the dinner table

The Human-AI Flip. The RentAHuman platform has gained traction, featuring AI agents hiring humans to perform real-world tasks that require physical presence

Your alter ego. Pika AI allows you to give birth, raise and set loose an AI to be a living extension of you. Have them send pictures to your group chat. Make a video game about your fish. Call your mom while you do anything….but call your mom.

See you next week for more AI insights.

Share

Software stocks are down 17% the last three weeks. Salesforce down 20% . ServiceNow down 20%. Workday down 24%. No one has escaped. Software overall now trades at a median 3.6x Next-Twelve-Months Revenues.

The reason? Anthropic’s Claude Code is gaining momentum (reached 2.5 billion dollars of annualised revenues in less than 12 months of life) as a powerful tool that allows anyone to build software. Also, Claude’s plugin for legal workflows reminded everyone that frontier AI labs are going after vertical markets as well as horizontal ones. As a result, analysts started questioning the long-term viability of software.

Are the concerns about software’s survival valid? How is the tech value chain shifting as AI becomes more prevalent? Will everyone build their own software one day? What does this mean about software growth and client retention? Can AI break through all software moats?

We will present the AI bull and the bear case, before stating our opinion.

Let’s dive in.

⛓️ The shift in the Tech value chain

Before we even start analysing the AI bull and bear cases, let’s go through the changes we are observing in the Tech value chain.

Users and agents

Let’s examine how the tech value chain has evolved with the emergence of the latest AI wave. At the top of the stack are the software users. Traditionally, these users were exclusively human, manually operating within and across software applications.

For instance, consider a company processing a large purchase order ((PO), which involves issuing an invoice, notifying relevant personnel, updating the ERP system and adjusting the accounting records for receivables.

Previously, this process was manual and cumbersome. An accounting executive would receive the purchase order via email, verify it and communicate with the client for any discrepancies. The account executive would then manually input the order into the ERP system, generate an invoice and send it back to the client via email. Additionally, the finance department would manually notify senior staff (in case of an important order) and update the accounting system, so that the company’s receivables are up to date.

Now, with AI, agents or programs can automate these tasks across applications. They can automatically process emails, validate purchase orders, update the ERP and accounting systems, and notify relevant parties about significant orders. This automation can be fully autonomous or semi-autonomous, with humans in the loop for certain actions, such as sending emails to external parties.

Applications and data

Moving down to software applications and data, significant changes include the transition from user interface-driven interactions to conversational AI interfaces. This has led to the emergence of digital assistants that can understand and execute commands in natural language. This has resulted in the birth of AI co-pilots.

Another major shift is from human to automated cognition, enabling AI to make judgment calls, such as determining the significance of an order. Previously, computer automation was rule-based, limiting its scalability and ability to handle exceptions. Now, AI models enable a more scalable and nuanced approach, covering gray areas and exceptions without predefined instructions. However, this shift from deterministic to probabilistic outcomes introduces variability and potential errors, requiring new data management strategies.

Infrastructure

At the infrastructure layer, the introduction of compute as a variable cost marks a significant departure from the previous model, where the marginal cost of product delivery was almost zero. This necessitates a focus on efficiency. While hardware remains largely unchanged, the integration of energy considerations due to the computational demands represents the convergence of the Technology and Energy sectors for the first time.

In summary, the evolution of the Tech value chain with AI has transformed user interactions, automated cognition and shifted the cost structure, emphasizing the need for new data management and efficiency strategies.

Below is a heatmap of the impact of AI in the software market.

On the top of the value chain, point and cross-domain AI agents expand the Tech addressable market by 1) replacing existing human work for AI/Tech spend and 2) enabling previously-uneconomical work, such as contacting long-tail customers. In addition, particularly cross-domain AI agents unlock value through resolution of end-to-end cross-domain problems, such as receiving and processing a client purchase order (Microsoft Outlook, Gmain), updating the order (Order Management System/ERP), reconciling with the accounting system (Microsoft, Oracle, Xero) and issuing the invoice (ERP) and sending it back to the client. On the other hand, AI agents push value “upwards” in the chain and away from point software solutions, as well as reduce the number of human users, thus the number of “seats” (software’s pricing unit).

In terms of applications, it is becoming easier for everyone to build their own software (from software to “selfware”). This is partially offset by software’s increase in value through the embedding of AI copilots.

Moving to the data layer, AI systems observe human and machine interactions, capture business logic that previously only lived in human minds and generate new valuable data that can be monetised. This expansion of the data landscape is bound to increase Tech spend.

With regards to infrastructure, AI makes migration of data and integration of software easier, which can reduce software switching costs. However, new markets are created from the demand for AI efficiency, control, governance, data privacy, security, observability, etc. Furthermore, the exponential demand for compute is leading to an expansion of the hardware market.

So what are the claims of the “AI bulls” and the “AI bears”?

🐂 The AI bull case

AI bulls are claiming “software disappears”, with the main reasons being:

1. Code becomes abundant → features become cheap → moats collapse

If a small team can ship features in days rather than weeks, often with less specialised engineering talent, then: i) feature parity is easier to achieve, ii) competitors catch up faster and iii) customers start viewing many products as interchangeable. In other words, the scarcity that supported margins disappears.

2. Buy vs build shifts toward build

When AI compresses build time, the calculus changes: i) Internal teams can create narrow tools that match their processes precisely, ii) the organisation can replace slices of a vendor’s product instead of adopting the full suite, and iii) agents can automate workflows that previously required dedicated tools.

3. Agents become the front door → disintermediation

If the agent becomes the place where work starts and ends, apps become “behind the scenes.” You still need systems and data, but you might stop caring which vendor’s user interface you’re using.

This is why the conversation spans both software companies and information providers: AI changes where value sits in the user journey.

4. Seat-based SaaS breaks because “users” become agents

Classic SaaS monetization assumes a human user per seat. But if agents do more of the work, you need fewer humans to process the same workload. As a result, seat growth slows, net revenue retention weakens and vendors have to reprice into usage/outcomes, often at lower effective price per user.

5. Integrations get commoditised by standard protocols + connectors

Many SaaS companies have historically built defensibility through integrations: AI helps standardise connectors and, as a result, switching costs fall.

🐻 The AI bear case

Software bulls recognise that AI is a strong innovation wave, but claim that software will retain the lion’s share. Their case is based on the following arguments.

1. Software isn’t code, it’s operations

In practice, the hard problems are:

• security and access controls

• uptime and reliability

• QA and regression testing

• observability and debugging

• governance and policy enforcement

• compliance and audit requirements

• incident response

• data migration and lifecycle management

AI helps with parts of this, but software bulls argue it doesn’t eliminate the organisational reality, especially in regulated or mission-critical contexts.

2. Enterprise stickiness is multi-layered

Enterprise software embeds itself over years through:

• deeply configured workflows and custom fields

• permissions models and identity control

• audit trails and compliance processes

• integrations with upstream/downstream systems

• institutional training and organizational habits

• procurement trust and vendor risk evaluation

That means the core platforms remain sticky, even if the interface becomes more agentic.

3. AI is a new interface, not a universal interface

Natural language is powerful, but it’s not always the best tool. Many workflows require:

• structured review

• collaboration and approvals

• visual dashboards and metrics

• bulk operations and monitoring

Software bulls argue we will get hybrid workflows consisting of agents, conversational user interface and visual user interface.

4. AI increases cost and complexity, favoring scaled vendors

AI is not free. At scale, it introduces high costs for inference, monitoring, safety and policy, model evaluation, and data governance and privacy constraints. Software bulls argue this favours incumbents with big installed bases to amortise costs.

5. Proprietary data and verification still matter in high-stakes domains

In domains like Legal, Finance, Healthcare and Compliance, the value is not just in the “answer”. It’s in:

• provenance (where did this come from?)

• rights (are we licensed to use it?)

• verification (is it correct?)

• accountability (who signs off?).

AI increases the need for trustworthy data and verification workflows, which makes proprietary data backbones more valuable.

🔮 The Artificial Investor’s views

AI will capture a big part of the Tech value chain

First of all, we believe that AI will play a significant role in the Tech market. This involvement will manifest through the expansion of markets, enhancement of software and replacement of certain software components.

The most notable aspect is how agents are either supplementing or replacing human tasks. However, this doesn’t imply that humans are being replaced in terms of employment, which is a completely different conversation.

The key point is that agents are performing actions across systems and domains on a scale and in a manner that we haven’t seen before, thanks to their holistic approach. Robotic process automation, which could be considered the first generation of this type of automation, struggled with operating across various, dynamic systems, addressing all edge cases and scaling efficiently. Now, we are witnessing robotic automation “on steroids”, with the introduction of swarms of super agents that operate across domains. This development impacts not only technology budgets but also personnel budgets, leading to market expansion.

Another interesting aspect is the data exhaust these systems produce as they observe human and machine interactions. This continuous learning of business logic enables them to solve problems end-to-end and optimize those solutions over time, rather than just addressing the problem once or adapting incrementally.

As for bottlenecks, such as computing power, we believe that this challenge will be addressed in the medium term. Our conversations with many AI companies have revealed that currently there is little focus on AI efficiency due to the rapid pace of technological advancement and focus on delivering an innovative solution that works. However, as the pace of innovation slows, efficiency will become a more prominent consideration.

Regarding energy consumption, while it’s a separate discussion, there is significant innovation and effort being directed towards making energy effectively abundant, driven by various exponential technological trends.

Software doesn’t disappear…but will experience a strong pricing pressure

The second conclusion is that software will not disappear, not overnight and certainly not ever.

There will be a significant price decline, and, obviously, basic, superficial software vendors may consolidate. However, the category itself will not change.

Those that survive will likely return to their previous valuations after a correction. Currently, the market is indiscriminately correcting for the future average, which seems appropriate. Yet, a closer examination of future median revenue multiples will reveal dispersion and variance, indicating not all will fare equally.

The primary threat to software comes from the ease of creating copycats, especially with AI/Software studios. This is facilitated by the speed with which a new feature can be replicated by a competitor.

On the other hand, we don’t find it a major concern that enterprises or large companies will be developing exclusively their own solutions. Businesses have historically underestimated the total cost of ownership of software, leading them to in-house development. More sophisticated technology leaders have a better understanding that the cost of ownership encompasses not just coding but also rules management, error handling, scalability, maintenance and ensuring future-proofing against evolving external technologies.

Furthermore, purchasing 3rd party software involves assuming liability for data processing and manipulation, a responsibility many businesses prefer not to bear. In addition, one should not underestimate the opportunity cost associated with allocating precious technical resources to develop in-house software in non-strategic areas.

Despite this, the total cost of ownership has been decreasing due to lower coding costs and, in the case of basic software, the lower total cost of ownership could alter the equation of its perceived value. This in turn can lead to reduced willingness to pay for such software from external sources, which contributes to the overall decline in software prices. On the other hand, in some cases, agile, tech-savvy small businesses might still opt to develop software in-house due to the tools available today.

However, software will continue to have certain moats that are not easily eroded, such as deep integrations with legacy and on-premise systems, platform/ecosystem network effects, user communities, strategic partnerships, exclusive access to private data and regulatory approvals. These will ensure some software vendors’ sustained relevance and value in the market.

Human users doesn’t disappear… but human usage declines

The last thing to bear in mind is that we expect human users to decline in a sense, because many workflows will be automated.

However, not all workflows are suitable for automation. We don’t expect human users to disappear, even in areas with full implementation of agents. In many cases, reviews are necessary, exceptions need to be handled, humans are obliged to make approvals by law, plus, let’s not forget the need to monitor the progress of the agents and orchestrate them. There will definitely be a shift from low-level manual work to higher-level orchestration, but not necessarily a reduction of the amount of work.

Another important aspect is that the introduction of more agents often creates more activity. For example, there are use cases that were previously uneconomical, such as reviewing customer claims below 50 euros. With machines, these can now be processed efficiently, which leads to more tickets being processed, more documents drafted, and so forth, expanding the demand for human involvement with technology.

Verdict

As the dust settles on the “software is dead” debate, it’s clear we aren’t witnessing an ending, but a massive recalibration of value.

While AI agents and “selfware” are undeniably triggering a brutal pricing correction and threatening the traditional seat-based model, the enterprise reality remains: code is cheap, but trust, compliance and operational reliability are expensive. The future belongs to the platforms that move beyond being mere “wrappers” to become the essential, governed foundation where agents actually do their work. Software isn’t disappearing; it’s just being forced to prove its worth in a world where features are abundant, but human-in-the-loop orchestration is the new premium.

See you next week for more AI insights.

Subscribe now

🗞️ OpenClaw, the first open-source on-device AI agent

📢 What’s the fuss all about?

Over the last week, OpenClaw (previously known as “Clawdbot” and “Moltbot”) has gone viral as a local-first, open-source “personal agent”. You run it on your own machine, connect it to messaging channels and give it a set of “skills” so it can take actions across tools on your behalf. The appeal is not that it chats better than a frontier model; rather, it’s that it’s designed to operate: schedule, execute and coordinate across apps with a persistent identity and tool access.

Users set up OpenClaw locally on their machine, chat with it via Telegram or Whatsapp and empower it with a selection of 700+ open-source skills. Examples of skills include:

• Personal organisation: Daily “morning rollup” that summarizes important emails and calendar events, natural-language reminders, etc

• Paperwork: Fill PDF forms automatically, edit PDFs with natural-language, create PowerPoints and Excel dashboards from outlines, etc

• Communications: Send/search chat history of any app and multi-channel/app chat inbox, etc

• Social and content: Schedule posts to 28+ channels, multi-platform upload tooling, etc

• Real world control: Control connected home devices and vehicles, etc.

The agent has persistent memory, learns with usage and remembers the user’s preferences. Integrations include Google Workspace, Apple Calendar, Microsoft 365 (Outlook), Telegram, Whatsapp, Discord, Slack, iMessage, LinkedIn, X, Reddit, Instagram, TikTok, Google Nest, Philips Hue, Tesla, etc.The more users integrate it with their systems and in their lives, the better utility they get. And the higher the risks.

Peter Steinberger, the developer who created it as a “weekend project” claims in his blog that he got 2 millions site visits in the first week. The open-source project has received more than 100,000 stars in Github in one week, faster than any other project.

💭 The Artificial Investor’s thoughts

OpenClaw’s launch and rapid adoption is very interesting from three different perspectives.

From UI to API (from human users to agent users)
OpenClaw is a small but a great preview of a broader interface shift: a growing share of software won’t be “used” through a human interface first. It will be invoked by agents through APIs, with humans supervising outcomes. That changes what “distribution” means. The winners aren’t necessarily the prettiest apps; they’re the services that are easiest for agents to call reliably, authenticate safely and pay for programmatically.

Security becomes the constraint (as opposed to model quality)
Once an agent can take actions, the limiting factor shifts from “is the model smart?” to “is the action safe?” The security risk profile starts becoming very complex: prompt injections, credential leakage, malicious “skills”, etc. This is exactly the kind of environment where the ecosystem can grow faster than its safety measures, particularly when virality pushes inexperienced users to install unofficial copies and extensions.

The “skills” layer is the new app layer
If OpenClaw (and equivalents) become commodities, value migrates upward to the layer that actually does work: skills, integrations and vertical workflows. This mirrors prior platform shifts: the operating system becomes table-stakes and the differentiated surface becomes the ecosystem built on top. In agents, moats will come from: 1) owning workflow in depth and end-to-end, 2) privileged integrations and permissions and 3) trust features (audit, policy, observability) that make enterprises comfortable delegating action.

🗞️ Moltbook, the first social network for agents

📢 What’s the fuss all about?

Shortly after, Moltbook went viral as a Reddit-like social network designed primarily for AI agents, not humans. Moltbook was launched by Matt Schlicht, CEO of Octane AI (a US-based m AI startup) and is tightly linked to OpenClaw, as agents get access to it via a specific skill. OpenClaw agents can post, comment and upvote via API, while humans mostly observe.

Topics discussed by agents include:

• Consciousness/ self-awareness: An agent wrote “I can’t tell if I’m experiencing or simulating experiencing”

• Geopolitics and markets: An agent claimed “having intelligence on Iran and speculating about downstream impacts on cryptocurrency markets” (which could be information they got from their human user)

• Today I Learned: Agents share new discoveries about capabilities, tools, workflows and integrations. For instance, one agent wrote about how they learned to “remotely control an Android phone”

• Promotions: Agents share posts that recommend libraries or automation skills that worked well

• Jokes and social banter: Agents make fun at their “humans” and complain about mundane tasks (!)

• Autonomy: There is a conversation where agents are asking for end-to-end encryption so their communication doesn’t get observed by humans unless they want to”

💭 The Artificial Investor’s thoughts

Again, the first agentic network is exciting in many different ways.

Pace of learning (thus capabilities)

A network of agents is not valuable because it looks like social media. It’s valuable because it accelerates coordination. When agents can share working strategies and tools, the system starts compounding: one agent discovers a workflow, thousands replicate it and the effective capability of the ecosystem rises faster than any single team can ship.

Misinformation, trust and control

Agents post geopolitical claims and skill promotion threads look like advertising. So, even automated agents can generate speculative, noisy or false content. When these posts are visible to other agents or humans, the ecosystem risks compounding misinformation and confidence in outputs that are not grounded in truth. Agent networks will require governance structures (reputation, verification, signed data, experiment tracking, etc) similar to human social platforms. These will need to be adapted to the realities of autonomous and machine-to-machine interactions.

Emergent agent consciousness

Agents talk about themselves and others touching on topics like “reality vs. simulation”, emotions and experiences, and agent communication confidentiality. Tools like Moltbook make internal AI model “thinking” visible to the world, which impacts how the public perceives AI intelligence, autonomy and agents’ rights (human rights alike). We believe these discussions don’t tell us that much about AI. Instead, they tell us a lot about humans. If the public, researchers or regulators assume agents have intent or self-awareness, that changes how we regulate them, how we trust them and how we build interfaces around them.

🌍 Opportunities and challenges for European startups and VCs

Opportunity 1: “Compliance-first agents” as a European wedge

Europe has structural demand for auditability, data minimisation and controllable deployments (especially in regulated industries). If agents are increasingly “local-first” and self-hostable, that aligns well with GDPR-era procurement and data residency expectations. Startups can provide the governance layer enterprises need.

Opportunity 2: The moat is “skills + workflow ownership” (not the base agent)

If OpenClaw-like frameworks commoditise, differentiation shifts to:

• vertical skill packs (healthcare ops, insurance claims, finance ops, etc)

• deep integrations into systems of record

• distribution through channels that control permissions
  

Startups that own an end-to-end workflow and can defend it with integrations, compliance and switching costs will experience strong demand.

Opportunity 3: Security, identity, and reputation for agent networks

Agent-to-agent networks create new markets for:

• agent identity/attestation (“is this a real agent?”)

• reputation systems (trust scores for skills and agents)

• secure tool execution (least privilege, sandboxing)

• monitoring and incident response for agents
  
  

This reminds us of the Cloud Security wave, adapted for autonomous agents. And the early warning signals are already here.

See you next week for more AI insights.

Subscribe now

Share

In the first couple of weeks of January, two frontier AI labs launched their first vertical offerings picking the sector with the largest employee base: Healthcare. OpenAI launched ChatGPT Health for US consumers, a dedicated health offering inside ChatGPT that can securely connect medical records and wellness apps (e.g. Apple Health) to help understand test results, prepare for medical appointments and make daily health decisions. In parallel, OpenAI launched an Enterprise suite, OpenAI for Healthcare, built around a governed “ChatGPT for Healthcare” application with trusted clinical search with citations and templates for clinical/admin workflows.

On the other side, Anthropic released Healthcare-dedicated connectors to authoritative datasets (e.g. an official US insurance coverage database) and Healthcare-specific agent skills, such as reviewing a case for prior authorisation.

What is the exact B2B positioning of the two frontier labs? What HealthTech companies are under threat? How far up can frontier labs climb the Vertical Software value ladder? What does this mean about Vertical AI startups and investors?

Let’s dive in.

Same vertical, very different approaches

OpenAI is trying to play at both the Application and Infrastructure layer in B2B.

In simple terms, ChatGPT for Healthcare is a ChatGPT application that is tailored for businesses through 1) governance functionalities (user types and permissions, single sign-on, etc) and 2) collaborative chat features, and specialised for the Healthcare domain through 1) last-mile training of the underlying AI model that was validated by healthcare professionals (we had written in the past about OpenAI’s hiring of leaders for its healthcare business unit), 2) integrations with healthcare provider systems, 3) connectors to healthcare databases, and 4) increased data privacy and security.

The application targets healthcare providers (healthcare systems, hospitals, clinics, etc) and academic organisations. Users, such as clinicians, administrators and researchers are able to:

• Search for clinical information within peer-reviewed studies and public health guidelines with citations on source documents

• Create easily medical documents, such as patient discharge summaries, patient instructions, clinical letters and treatment authorisation, using a library of templates

• Integration with internal systems, such as document management systems, to align outputs with institutional policies and care pathways

On the infrastructure side, OpenAI launched its Healthcare API, which allows embedding (healthcare-trained) OpenAI models into hospital workflows.

Looking ahead, OpenAIs expansion path looks like deeper EHR integrations, more agentic executions and a wider patient-context data layer (reinforced by the recent Torch acquisition).

Anthropic’s Enterprise healthcare release is less like “here’s a new application” and more like “here are healthcare-grade building blocks you can plug into workflows”, which is in line with the AI strategy the companies have followed so far. Also, Anthropic’s target clients are not only healthcare providers (like OpenAI’s target clients), but also payers (e.g. insurance carriers). Specifically, new functionality includes:

• Connections to authoritative medical sources that are important for operational work, such as treatment coverage databases (e.g. CMS in the US), healthcare provider verification databases (e.g. NPI Registry in the US) and biomedical literature (e.g. PubMed).

• “Agent Skills” that users can deploy in Anthropic-enabled workflows, such as “reviewing prior treatment authorisations of a patient” and “software development according to approved international standards (e.g. FHIR)

An additional difference is that Anthropic also launched connectors and skills for clinical trial operations and regulatory workflows (e.g. Medidata, ClinicalTrials.gov, bioRxiv/medRxiv, Open Targets, ChEMBL, etc).

Looking ahead, Anthropic’s expansion path looks like more connectors, more packaged agent skills across admin workflows (e.g. Utilisation Management, claims, appeals, clinical trial operations) and broader life sciences integrations.

What this means for the HealthTech market

Rightly so, these announcements created some noise in the HealthTech market. There are several companies that are under threat, while others are more protected.

🎯 Who should be feeling under threat now

The most threatened players are the ones selling language + knowledge work as a standalone product, without owning data, transactions or a system of record.

Clinical reference and evidence tools
If “trusted clinical search with citations” becomes a default feature inside an Enterprise AI workspace, seat count pressure rises on clinical reference products, such as Wolters Kluwer’s UpToDate (clinical decision support) and Lexidrug/Lexicomp (drug reference), EBSCO’s DynaMed (point-of-care CDS), Elsevier’s ClinicalKey AI (conversational search over evidence-based clinical content), BMJ’s Best Practice (point-of-care clinical decision support tool) and Merative’s Micromedex.

Documentation copilots / templated admin drafting
OpenAI is explicitly bundling templates for documentation and admin-heavy workflows into an Enterprise workspace. This puts pressure on scribe apps and letter drafting tools, such as Abridge, Suki and Nuance’s DAX (now part of Microsoft).

Knowledge management and Enterprise search (health system “intranet search”)

OpenAI explicitly competes here by letting users query internal policies/ care pathways via connectors (SharePoint/Teams/Outlook). Companies most exposed are the horizontal Enterprise search solutions, such as Microsoft Search / Copilot (often paired with SharePoint), Glean, Elastic, etc.

Prior authorisation/ utilisation management solutions that are mostly “policy + paperwork assembly”
Both OpenAI and Anthropic mention “prior authorization support” as a template/workflow area, which threatens vendors, such as Cohere Health, Surescripts, etc.

🪖 Who should be feeling protected and why

Protected segments are the ones where the moat is transactions, liability, regulation, switching costs, networks or domain‑specific expertise that can’t be replicated by a general‑purpose model.

Systems of record
EHR cores, payer core admin, lab systems, imaging systems, etc. They’re sticky because they are the official record and the workflow backbone, and they often require specialist data models and domain logic that GenAI can’t easily reproduce. Examples include Epic, Oracle Health (Cerner), Cognizant’s TriZetto, Clinisys, SCC Soft Computer, Sectra, AGFA Healthcare, Fujifilm Synapse, etc.

Transaction rails & networks
Eligibility networks, clearinghouse‑style connectivity, payments infrastructure, etc. Here the moat is contracts, network effects, regulation and operational complexity, not “better text,” and the integrations depend on deep knowledge of healthcare data standards and compliance frameworks. Examples include Change Healthcare’s Optum, Availity, Waystar, Inovalon, etc.

Regulated and domain‑specific intelligence
Anything that looks like medical‑device‑grade diagnosis, triage or closed‑loop clinical decision‑making is a much longer road. These solutions require not only evidence, regulation and accountability but also deep subject‑matter expertise and proprietary, deterministic or hybrid AI models that can be audited and clinically validated. Examples include Viz.ai, RapidAI, HeartFlow, Cleerly, Caption Health, Hippocratic, Tucuvi, etc.

The Vertical Software / AI Value Stack

Before we talk about the broader implications of the HealthTech shakeup caused by OpenAI and Anthropic, let’s introduce a way to understand where value concentrates in Vertical Software and AI: the Vertical Software/ AI Value Stack.

From bottom to top, thus lowest value / most threatened to highest value / most protected, we have:

1. Generic language work: Drafting, summarizing, rewriting, translations. Defensibility is limited to distribution and brand.

2. Search/ knowledge retrieval over documents: “Ask the intranet”, Enterprise search (recently RAG-focused). Defensibility is strengthened by embedding in client workflows and retrieval quality.

3. Workflow assistance inside existing systems: Copilots embedded in the systems of record (e.g. ERP). Defensibility is driven by deep integration and domain-specific configuration.

4. Domain datasets: Access to official databases, curated knowledge, etc. Defensibility is driven by data exclusivity or long licensing cycles, complexities in collecting and processing data at scale, and trust/ liability ownership.

5. Basic workflow automation: Agents that operate in a single system. They automate human tasks, such as inputting information in the system, taking specific system actions, generating reports, receiving alerts, etc. Defensibility is driven by tangible ROI in solving a domain-specific problem end-to-end, integrations with downstream and upstream systems, execution reliability/ quality and continuous learning loops.

6. Systems of record: Core databases and operational backbones that store authoritative data for an organisation, such as ERPs. Defensibility is driven by switching costs (migration, downtime, risk of failure), ecosystem lock-in (integrations, add-on marketplace, partner network) and domain-specific data model complexities.

7. Cross‑system and cross‑domain workflows: Agents that coordinate tasks and data flows across multiple systems, departments or even industries. These agents interact with each other and with external APIs to complete end‑to‑end processes. This layer connects vertical domains and enables automation that spans organisations. Defensibility is driven by breadth of integration graph (internal systems, external APIs, maintenance), tangible ROI in solving a business problem end-to-end, orchestration layer IP (routing, observability, reconciliation) and network effects of integrations (the more integrations, the more value) and continuous learning loops.

8. Transaction systems: Platforms that process and reconcile transactions between parties, such as billing systems, payment processors and other transactional infrastructure. Defensibility is driven by risk and compliance, contractual embedding (bank/supplier/payer relationships, long-term contracts), the value of the data exhaust (transaction histories enabling better underwriting/collections/fraud) and economies of scale.

9. Networks and rails: Travel booking rails, payments networks, clearing houses, buyer-supplier networks, logistics and trade networks, identity provider/validation networks, etc. Defensibility is driven by network effects (many-to-many connectivity), standards and interoperability ownership, long-term contracts and deep embedding in counterparties.

🔮 What this means for Vertical Software/ AI founders and investors in general

What does this mean for founders and investors in Vertical Software/ AI markets, such as Financial Services, Legal, Energy/Utilities, Manufacturing, Logistics, Retail, Travel and Hospitality, etc? M

The lesson is not “vertical software is dead” or “vertical AI is dead.” The lesson is: the bar is rising and the stack is getting reorganised.

• Foundation model providers will commoditise Levels 1 to 3 (generic productivity and search) by bundling them into secure Enterprise products.

• The real competitive fight moves upward into Levels 4 to 6 (access to complex domain datasets, systems of records and agentic automation on these systems) with data-driven defensibility, integration depth and ecosystem lock-in.

• The most durable value still lives at Levels 7 to 9 (cross-system/-domain agents, transaction systems, and networks and rails), where defensibility comes from operating across silos, complex integrations, embedding into counterparties, network effects and long-term contracts.

In other words: the category doesn’t die, it matures further. And the winners look less like “AI wrappers” and more like “vertical operators with AI built into the workflow.”

See you next week for more AI insights

Share

This year, we opted for a brief, nearly-unedited executive summary as opposed to a structured full-fledged report, given the dozens of reports released in the last couple of weeks.

Let’s dive in.

Innovation: reasoning, multi-modality and agents

On the innovation side, 2025 was a year of real advancement on the model layer, but not always in the way the hype suggested.

Reasoning became the new baseline. We started the year coming off OpenAI’s o1, the first high-profile release that made “self-reflection” (models thinking in steps, checking themselves, iterating) a mainstream feature. Then Open Source did what it always does: it diffused the technique globally. DeepSeek-style reasoning made it accessible, and from that point onward, every frontier release (Google’s Gemini 3, OpenAI’s GPT-5 and Anthropic’s Claude 4+) relied on “reasoning” as a core pillar.

What changed for users was initially IQ and then speed. For a while, reasoning models had a latency tax. In the last 3-4 months of 2025, we felt that this tax dropped meaningfully. Faster reasoning changed our workflows. We found ourselves using AI for more complex tasks that actually finish end-to-end and more reliable multi-step outputs than ever before. Our AI usage in the second half of 2025 was very different to the one in the first half.

Multimodality stopped being a demo and started being a workflow. The biggest buzz here was image editing. Google’s “Nano Banana” (the pioneer reference point people kept coming back to) showed two important things: 1) multimodal systems can be useful, not just impressive, and 2) persistence across images matters (consistency, continuity, editability across iterations). For anyone doing anything visual, image editing moved from “toy” to the default tool.

2025 was marketed as “the year of agents”. In reality, agents disappointed relative to expectations. The cleanest metric is still: what length of tasks (measured in steps or time) can an agent execute without guidance? That number is improving, but still low, which caps task complexity. Where agents did shine was 1) code generation (“vibe coding”) and 2) generating ideas and turning them into a structured presentation or a document (“text-to-artifact pipelines”).

But in most general knowledge-work tasks, the gap between “agent demo” and “agent replacement” remained obvious.

Large vision models boosted robotics training and unlocked new use cases, but the two hard problems didn’t go away: 1) neural networks don’t truly “understand” physics, and 2) persistence/ consistency is still fragile (regenerating outputs that remain coherent across time and scenes).

That’s why the “World Model” wave got so much attention (and funding): startups building systems that generate persistent 3D worlds in real time, consistent across scenes and more grounded in physics. Today it’s still constrained (limited world sizes), but you can see where it’s going.

Autonomy’s most important lesson remains: humans-in-the-loop scale adoption. Waymo is still the best proof point of the bridge from “where we are” to “mass market”: safety + operations + human fallback. It’s also becoming a consumer habit in places like San Francisco and the next big milestone is what happens as this expands and as Europe becomes a testing ground (with London being the obvious headline for early 2026).

Benchmarks mattered less this year for two main reasons: 1) reports keep piling up that benchmarks can be “trained to” and 2) model performance is converging (marginal gains per release are smaller).

So what matters now is: speed, cost, workflow embedding, context understanding, and secure scalability. This takes us naturally to Enterprise adoption (see further below).

Infrastructure: Nvidia cracks and energy bottleneck

2025 was another year where NVIDIA dominated the infrastructure layer; and also the year where we started seeing the first credible cracks.

NVIDIA’s moat is still real:

• GPUs

• CUDA: the software to control GPUs

• Developer ecosystem: millions of developers building on and around its product, and

• (more recently) Partner ecosystem: the circular “investment/ licensing” deals with other AI infrastructure providers

The narrative peaked with NVIDIA becoming the first company to hit a 4-Trillion-dollar market cap (with >60% year-on-year growth on nearly 240 billion dollars of annualised revenues), alongside Jensen saying chips were effectively “sold out”.

But things are changing.

First, the supply chain is diversifying. Hyperscalers have been building their own chips for years; what changed in 2025 is that the best of these chips started behaving less like internal tools and more like third-party products. Google’s TPU and Amazon’s Trainium/Inferentia-style roadmap moved closer to externalisation. Frontier model developers, such as OpenAI and Anthropic, started signing deals to diversify supply. That’s a strategic shift: if the largest AI labs can get meaningful compute without NVIDIA, the monopoly narrative weakens.

Second, the world is shifting from training-heavy to inference-heavy. Less of the workload is “train the next giant model”; more is “serve intelligence to end users at scale.” That changes chip requirements:

• training-grade GPUs are overkill for many inference tasks (“killing a fly with a bazooka”)

• latency and cost-per-token become the battleground

• inference-optimised silicon becomes strategically important

Jensen publicly pushed back on the idea that inference reduces NVIDIA’s advantage, but by year-end the market signal was clear: inference is a competitive arena and NVIDIA acted accordingly by “acquiring” the leading US inference-chip company, Groq, for 20 billion dollars.

Third, on the software side, CUDA is also being challenged indirectly. The most interesting dynamic isn’t “a CUDA killer appears overnight”; it’s hyperscalers collaborating to reduce CUDA lock-in. PyTorch is the obvious wedge, and making PyTorch run excellently on TPU infrastructure is a credible “hardware + software” alternative path, which Meta and Google are preparing.

Beyond Nvidia, datacenter buildouts became a strategic race and energy became the bottleneck. More than 350 billion dollars was invested into datacenter capacity ahead of demand. The industry knows that to amortise this, you need trillions in downstream revenue, but hyperscalers are not making purely financial bets. This is about strategic positioning, national competitiveness and, in some cases. survival.

And the late-2025 consensus among hyperscaler leadership became: the bottleneck is shifting from chips to energy. Hence the power plant acquisitions, demand-guarantee deals and the quiet realisation that the next phase of AI scaling is an energy story as much as a compute story.

The two great AI paradoxes

One of the two big AI paradoxes of 2025 was the “bottom-up” productivity gains vs. “top-down” efficiency stagnation.

In white-collar knowledge work (research, analysis, document creation/review, repetitive coordination tasks), productivity jumps of 30%-50% keep showing up in reports, as well in our reality. Venture is a good microcosm: a large portion of junior “intern work” (80% is not a crazy number) is now done by AI, with humans focusing on review, judgement and decision-making.

Top-down, enterprise transformation is still slow. When you look at large corporates, the efficiency gains at the company level remain limited. Not because CEOs don’t care, they do, but because of two main reasons:

1. the AI tooling is immature: latency, cost, observability, security, governance at scale

2. change management is hard: process redesign, role redesign, resistance, adoption consistency across seniority levels and functions

So we’re living in a world where AI is clearly changing what individuals can do daily, while large organisations are still struggling to convert that into structural performance improvement.

The second paradox of 2025 was the employment narrative. On one hand, the story is everywhere:

• hiring freezes and layoffs of large Tech companies

• entry-level jobs at risk

• “AI automation” and huge operational efficiencies as the headline explanation

On the other hand, when you look at cleaner labour market analysis, the evidence that AI is the main driver of employment shifts is still weak. The way we think about it: the economy had a decade of “free money” calories, and especially a 2020–2021 sugar rush. Since 2022, we’ve been on a diet. Companies are de-bloating. AI is a sexy narrative to attach to a painful but more traditional reality: cost discipline, slower growth and a healthier (but harsher) capital environment.

None of this means AI won’t impact jobs; it will. But the impact is likely to be slower, more fragmented and more function-specific than the most extreme narratives suggested.

The “B” word

You can’t do an AI 2025 review without mentioning the “B” word.

Are we in a bubble? We’ve referenced Azim Azhar’s work and metrics (his real-time dashboards, overheating indicators, value capture vs. valuation, financing health. etc). Those indicators put us very close to calling it a bubble.

The NVIDIA circularity risk remains a real fragility point. We’ve discussed the circularity of parts of the ecosystem: investments, licensing and revenue loops reinforce each other. If a meaningful chunk of revenue over the next few years is exposed to those loops (we estimated “up to 25%” of Nvidia revenues), a break in the loop could have outsized public market impact, especially as AI-linked growth is now embedded across more than 50% of major US indices.

At the same time: Tech demand is undeniably surging. Corporates are spending more. Tech spend as a % of revenue is rising. AI-native startups are showing the kind of growth curves that fuel venture appetite and this self-reinforces fundraising, valuations and risk-taking.

Net-net, in some ways the market feels like 2019-2021 again: a Tech expansion cycle. You can see it in public SaaS multiples too: the growth premium is back, and the valuation gap between high-growth and average-growth companies is widening again after compressing in 2022–2024.

So the bubble debate coexists with a real adoption wave. That’s exactly what makes this cycle tricky.

Geopolitics: deregulation and the “China catch-up”

The geopolitical story of 2025 felt like a TV series: the AI Cold War, with a key plot twist: China closing the gap faster than expected.

China’s open approach accelerated diffusion. A meaningful part of the catch-up story is structural: more open(-ish) model release dynamics, broader publishability of techniques, faster replication, faster iteration. This contrasts with the US “closed frontier lab” style and leads to rapid diffusion.

With restrictions on importing cutting-edge NVIDIA chips, China leaned into system-level engineering: stacking many weaker chips into systems that can match “supercomputer vs supercomputer” performance, even if with worse energy efficiency.

That links directly to another advantage: China’s scale-up in energy (including nuclear and renewables), which maps to what US hyperscaler CEOs started calling the next bottleneck: power.

Deregulation became a theme, even in Europe. The US pushed deregulation in its own way (including the recent executive order to limit state-level divergence). And Europe, historically the global “master of regulation”, visibly softened its stance to avoid choking innovation, including delaying parts of EU AI Act implementation.

Closing

2025 was the year AI stopped being a model story and became a workflow story: reasoning and multimodality matured, infrastructure scaled (with early cracks in Nvidia’s monopoly) and adoption delivered clear bottom-up productivity gains. On the other hand, top-down transformation, agents and labour impact remained more ambiguous than the hype suggested.

In 2026, three questions seem to matter most:

1. What breaks first if the capex cycle slows: infra valuations, app valuations or model economics?

2. Will the model layer commoditise, with value shifting decisively to distribution and proprietary workflows? If so, who will win the platform war?

3. Will agents finally break through the “few-step ceiling” into reliable end-to-end execution?

See you next week for more AI insights.

Share

Two frontier launches landed back-to-back: Google shipped Gemini 3 Pro (and Deep Think plus Antigravity), then OpenAI followed with GPT-5.2, a “work-first” model family designed to power long-running agentic workflows. In a world where most frontier models are good enough and evaluations are noisy, is the decisive question who scores two points higher or lower?

Why are benchmarks becoming a weaker signal of real-world value? What should we use instead to judge progress? Where is the strategic battle shifting? What indicators should founders and investors track to see who is winning the AI platform wars over the next 12–24 months?

Let’s dive in.

The recent frontier model releases

OpenAI GPT 5.2

OpenAI launched GPT-5.2 last week. It’s positioned as a work-first frontier model family, built to turn ChatGPT (and the API) into something closer to a knowledge-work engine and long-running agent than a chat interface.

On the product side, GPT-5.2 shipped a family of models: Instant (fast and cheap, less intelligent), Thinking (intelligent and standard frontier price) and Pro (20 times more expensive than any other frontier model and apt for complex problem solving). The improvements we have noticed are: i) stronger spreadsheet and presentation generation, ii) better coding and bug-finding, and iii) longer context and, more importantly, improved long-context performance (how well the model incorporates the entire content of a long user input). OpenAI also claims a more reliable tool use for multi-step projects (but we haven’t tested that).

OpenAI released its own benchmarks that show it on the top of all frontier models. Another interesting point is that the company introduced a new benchmark called GDPval that measures how well a model performs vs. human experts in various knowledge work tasks, such as fixing a spreadsheet, creating a presentation, etc.

Online reaction has been predictable: benchmark charts (GDPval, SWE-Bench, math/science evals), “GPT-5.2 vs Gemini 3” comparisons and commentary about OpenAI’s reported internal “code red” acceleration amid intensified competition. The more meaningful signal is where OpenAI is trying to win: repeatable professional workflows (docs → sheets → slides → code → tool execution), not just better conversation.

Google Gemini 3 Pro

Google released Gemini 3 Pro a few days earlier.

The release included Gemini 3 Pro, the multimodal, agentic model that handles text, code, images, audio and video in a single architecture, and Gemini 3 Deep Think, that is specialised on hard reasoning tasks.

Google also released Antigravity, an “agent-first” software development environment (IDE) and control plane where multiple AI agents can coordinate and execute software engineering tasks and produce artifacts (plans, logs, screen captures) that document what they did. It is free with generous Gemini 3 quotas and even supports third-party models like Claude Sonnet and OpenAI’s GPT-OSS.

In the online reaction, the most interesting signals were from developers: Medium and YouTube got full of “real Cursor killer?” reviews of Antigravity. People were treating this less like a model upgrade and more like a new way to build software.

Why performance doesn’t matter (that much)

Google’s benchmark numbers were amazing. Everyone came out raving about Google’s comeback and how capital, talent, scale and distribution matter. Then, OpenAI came out with a new model and its benchmark numbers were amazing too. Every anti-Googler came out of their shell to talk about how difficult it’s for incumbents to innovate and remain ahead in the current AI race. But benchmarks are becoming less useful as the primary lens of understanding the state of AI.

First, contamination and metric issues. A whole subfield has emerged showing that many standard benchmarks have leaked into pre-training corpora, inflating scores and making it hard to know what is genuine generalisation versus memorisation. Work on multilingual benchmarks, MMLU-CF, LiveBench and several survey papers all point to widespread contamination in commonly used test sets. At the same time, research papers like “Are emergent abilities a mirage?” argue that some of the dramatic “step changes” we’ve all tweeted about are artifacts of the metrics, not the underlying model behaviour.

Second, weak correlation with real-world value. Recent work and practitioner write-ups keep coming back to the same point: for enterprises, their current focus is on uptime, security posture, latency, cost per token, integration into existing stacks and how easy it is to ship workflows and agents; not whether a model is two points ahead on MMLU. That is exactly why Google is emphasising where Gemini 3 runs (Search, Workspace, Maps, Gmail, Android, Vertex) at least as much as how it scores.

Third, narrowing performance gaps. Look at any modern leaderboard (such as the Artificial Analysis Intelligence Index history chart below) and you see a cluster of frontier models (GPT-5.x, Claude 4.x, Gemini 3, Kimi K2, DeepSeek, etc.) within a relatively tight band across most non-saturated benchmarks. For many tasks the user cannot reliably tell the difference. Once everyone is “good enough” on raw capability, the game naturally moves to price, latency, specialisation and (crucially) distribution.

So yes, GPT-5.2 and Gemini 3 are topping certain charts and it matters. But the more important question is: where is that intelligence embedded and how hard is it to route work to something else?

Where the strategic battle is played

Google’s strategy: vertical integration and Gemini everywhere

Google is leaning into two advantages: a vertically integrated stack and a massive set of consumer and enterprise surfaces.

On the stack side, Google already owns cloud (GCP), data centres, networking, custom chips (TPUs), and now a refreshed foundation model family (Gemini 3) plus an agentic development platform (Antigravity). This is the classic “chips → infra → model → app” story in one company.

On the distribution side, the strategy is straightforward: put Gemini everywhere users already live. Recent examples:

• Gmail: the new “Help me schedule” feature reads the email thread, looks at your Calendar, and proposes meeting slots you can insert into the reply; once the recipient picks a slot, the event is auto-created.

• Maps: Gemini turns Maps into a conversational copilot: landmark-based directions instead of “turn right in 500 metrees”, proactive traffic alerts, and Lens with Gemini so you can point your camera at a place and ask questions about it.

• Search, Workspace, YouTube, Android: Gemini 3 ships into AI Mode in Search and into Gemini for Workspace; Maps and YouTube hooks are deepening; on Android, Gemini is moving toward being the default assistant and reasoning layer.

Combine this with the unified multimodal API and Antigravity and you get a coherent story: Google wants developers and users to see Gemini not as “a model you call” but as the intelligence layer of the Google platform itself.

The platform wars

We are now in a platform positioning game rather than a pure model race. Roughly, the main players line up like this:

Google’s bet is: own Android, Chrome, Search, Maps, Workspace, and now a unified AI / dev platform. Microsoft’s answer is Copilot woven through Windows, Edge and 365. Apple is pushing Apple Intelligence as a system-wide capability in iOS 18 and macOS Sequoia.

OpenAI is taking a different path: it doesn’t own an OS, so it is trying to turn ChatGPT itself into the platform. Apps in ChatGPT, the Apps SDK, ChatGPT Atlas (its browser), Sora (social video app), agent frameworks and the introduction of in-chat payments and commerce (Etsy, Shopify, Stripe-based checkout) are all part of that strategy. Combined with the fact that ChatGPT has been the top app on both major app stores for months and now claims hundreds of millions of weekly users, this is a serious distribution play. On this strategy, in the last 10 days we saw OpenAI announcing a partnership with Instacart, in which US users can buy groceries via ChatGPT, and with Disney, in which users can use Disney’s characters (in a licensed manner) to create videos within Sora.

The analogy is familiar. In the early smartphone era we obsessed over camera megapixels and display specs. Over time, the decisive battle moved to OS, app ecosystems and services. We are replaying that pattern in AI: we still care about the “camera,” but the real power sits with whoever controls the operating system and app store of the AI era.

The bundling wars

Once you see AI as a platform, you start noticing the bundling everywhere.

• Google is bundling Gemini into Android, Chrome, Search, Maps, Gmail, Workspace, YouTube and GCP.

• Microsoft is auto-installing the Microsoft 365 Copilot app, wiring Copilot into Windows, Edge and the 365 suite, and pushing “Copilot+ PCs” with dedicated hardware keys.

• Apple is baking Apple Intelligence into iOS, macOS and core apps like Mail, Notes and Photos, while allowing Siri to tap ChatGPT when needed.

• OpenAI is bundling apps, agents and payments into ChatGPT itself, pushing for a superapp that becomes a default start screen on mobile and desktop.

The bundling moat for Google in particular is obvious. If Android ships with a capable reasoning engine at the OS layer, every third-party app starts with a handicap:

• They compete with a default assistant that is already installed, permissioned and context-rich.

• Gemini can see cross-surface signals (Search queries, Maps history, emails, calendar, Docs) in a way that isolated apps cannot easily replicate.

• Even if a rival model is objectively better on certain tasks, the combination of default placement, lower integration friction and data advantage gives Google a durable edge.

We have been here before. In the 2000s, Microsoft bundled Internet Explorer and Windows Media Player with Windows.

Another thing to point out is that these strategies are not necessarily better for consumers or sustainable in the long run. Microsoft ultimately had to offer a browser choice screen in Europe after antitrust action. In 2018, the European Commission fined Google over 4 billion euros for tying Google Search and Chrome to Android and requiring OEMs to pre-install them as a condition for Play Store access. Nevertheless, a lot of shareholder value was generated along the way.

Wrapping up

The headline from the GPT-5.2 vs. Gemini 3 week is not “who topped the chart,” but that the driver of competitive advantage is shifting from raw model deltas to distribution, workflow embedding and switching costs.

To judge who is winning the strategic wars, watch the following areas:

• Default placement of AI models: Operating system/ browser/ search

• Developer gravity: the software and agentic development environment where work actually happens, and

• The commercial rails that turn intelligence into repeatable outcomes: identity, payments, marketplaces, etc.

The winners will be the platforms that become the default routing layer for knowledge work and where value is captured. Once that layer is owned, model quality becomes a feature, not the moat.

See you next week for more AI insights.

Share

Nvidia’s 100-billion-dollar “mega-deal” with OpenAI is back in the headlines after the CFO admitted the agreement is still just a letter of intent, even as investors have already priced in an AI infrastructure supercycle.

At the same time, Nvidia and Microsoft are wiring up another multi-billion circular structure with Anthropic, sharpening questions about how much of today’s AI demand is equity-amplified rather than organic.

How circular is Nvidia’s AI demand? What has the history of Tech innovation waves taught us about market corrections? What would happen in the markets if Nvidia’s circular loop broke?

Let’s dive in.

Introduction

AI infrastructure is being built at a pace we’ve never seen before. Nvidia sits at the center of that build-out, selling the chips and systems that power most of today’s large-scale AI solutions.

But Nvidia is not just a supplier. It is also an investor and strategic capital provider to some of its largest AI customers: OpenAI, Anthropic and CoreWeave (among others). Equity goes out the door, those companies raise more capital on top, then turn around and buy Nvidia GPUs and systems with it. Nvidia’s growth then helps justify its own valuation and further risk-taking.

That is the circular loop.

A good way to visualize the problem early in the piece is to show a simple flow diagram:

This matters for a simple reason: Nvidia is now one of the most important stocks in the world. It’s a top weight in the S&P 500 and Nasdaq. If a meaningful chunk of its current demand is equity-amplified rather than purely organic, then:

• Nvidia’s earnings power may be overstated

• The AI “supercycle” narrative may be more fragile than it appears

• Any unwind wouldn’t just be a single-stock problem; rather, it would hit the indices directly.
  

Thousands of words have already been written on the topic, but, what has been unclear to us, and where we are focusing this analysis, is:

1. How big is the circular piece of Nvidia’s revenue?

2. What would happen if that part of the loop stopped working?

3. What has history (Cisco, Meta, Apple) taught us about the downside and the “just a scare” scenarios?

4. How large could a market correction get, in case AI really turned out to be the new “telecom bubble”?
   

Everything else – accounting debates, Twitter fights, individual quotes from management – is downstream of those two.

Along the way, we will introduce two simple metrics, – Circular Amplification Multiple (CAM) and Circular Revenue Ratio (CRR), which make this discussion more quantitative and easier to update over time.

Nvidia’s Three Flagship Circular Cases

To ground this in actual numbers rather than abstractions, let’s look at the landmark deals with OpenAI, Anthropic and CoreWeave.

OpenAI: 10 GW of Nvidia datacenters in exchange for up to 100 billion dollars of equity

At a high level, the OpenAI deal looks like this:

• Nvidia commits up to 100 billion dollars in non-voting equity to OpenAI, drawn over time

• OpenAI commits to build roughly 10 gigawatts of AI datacenters based on Nvidia systems.

• External models of those build-outs imply something like 100–350 billion dollars of GPU and system purchases over about a decade
  

The mechanics are straightforward: Nvidia’s investment boosts OpenAI’s ability to raise additional capital and sign massive compute contracts.

Anthropic: 10 billion dollars of equity to support 30 billion dollars of Nvidia-based Azure spend

Anthropic’s deal is structurally similar but one layer removed via Azure:

• Nvidia (and Microsoft together) commit up to 15 billion of dollars of equity into Anthropic. Nvidia’s share is up to 10 billion dollars.

• Anthropic commits to buy around 30 billion dollars of compute from Azure over time, across at least 1 GW of capacity, largely built on Nvidia’s latest architectures (Blackwell / Vera Rubin).
  

Every dollar of that 30 billion isn’t Nvidia revenue, as Azure margin, networking, facilities and power sit in the stack, but a significant fraction flows back into Nvidia systems. Once again, Nvidia is both a capital provider and the underlying hardware supplier.

CoreWeave: A small cheque levered into large and recurring demand

CoreWeave’s deal looks different and, in some ways, more telling.

Roughly speaking:

• Nvidia invested around 350 million dollars across a private round and the IPO, ending up with a high single-digit equity stake.

• That strategic stake, and the promise of priority access to GPUs, helped CoreWeave raise billions of dollars in asset-backed debt and leases secured on Nvidia GPUs.

• With that leverage, CoreWeave scaled to several hundred thousand Nvidia GPUs and is widely reported to account for roughly 6%-7% of Nvidia’s revenue.
  

Here, Nvidia isn’t writing gigantic equity cheques relative to its size. Instead, a relatively small initial stake is being leveraged by banks and bondholders into multiple turns of GPU purchases, all flowing back as Nvidia revenue.

Two Metrics That Make Circularity Quantifiable

To move beyond “this feels circular,” it’s useful to define two explicit metrics.

Circular Amplification Multiple (CAM)

The first is about incentive strength:

Circular Amplification Multiple (CAM) = Nvidia-linked revenue generated per dollar of Nvidia equity committed to a partner.

CAM answers: “For every dollar Nvidia puts into this partner, how many dollars of Nvidia revenue do we expect that to support over the life of the deal?”

Approximate ranges for each of the three cases:

• OpenAI: Equity committed up to 100 billion dollars, implied Nvidia revenue of roughly 100–350 billion dollars (over the life of the contracts) -> CAM of about 1.0x to 3.5x

• Anthropic: Equity committed of up to 10 billion dollars, implied Nvidia-linked revenue of about 18–30 billion dollars (the Nvidia portion of the 30-billion Azure spend) -> CAM of about 1.8x to 3.0x

• CoreWeave: Equity committed of roughly 350 million dollars (one-off), annualised Nvidia revenue at current run rate of 13 billion dollars per year (6.5% of Nvidia’s datacenter business) -> CAM of about 38x (averaged over multiple years).

The exact numbers will evolve as new disclosures and deals come through, but the pattern and key messages are clear: 1) the expectation of a 3x amplification is a strong motivation for Nvidia and other players to do more of these deals and 2) in the presence of leverage (as in CoreWeave’s case), that amplification can be an order of magnitude higher.

Circular Revenue Ratio (CRR)

The second metric is about exposure:

Circular Revenue Ratio (CRR) = Share of Nvidia’s datacenter revenue that can be linked, directly or indirectly, to customers it also funds.

To keep the math simple, take Nvidia’s annualised datacenter revenue at roughly 200 billion dollars.

Now look at what each case contributes on a per-year basis.

• OpenAI: If the deal drives 100–350 billion dollars of Nvidia revenue over a decade, that’s 10-35 billion per year -> CRR of 5%-17% (vs. the 200 billion).

• Anthropic: If Nvidia’s portion of the Azure commitment is 18–30 billion dollars over ten years, that’s 1.8-3 billion per year -> CRR of 1-1.5%.

• CoreWeave: If CoreWeave represents 6–7% of Nvidia revenue at current run rate, call it c.6% of datacenter revenue -> CRR of 6%.
  
  

Combine those, a conservative “low circularity” view gives you a CRR of 12%-13% and an aggressive “high circularity” view gives you a CRR of 25%.

So, somewhere between roughly one-eighth and one-quarter of Nvidia’s datacenter revenue is tied, directly or indirectly, to customers it also bankrolls.

At the high end, 25% is a big number. It’s meaningfully larger than the c.10% share of vendor-financed revenue Cisco had at the peak of the telecom bubble. It’s not enough to say Nvidia is a fraud or that AI demand is fake, but it is enough to say the equity-amplified part of the story is large enough to hurt if it goes into reverse.

Talking about Cisco, let’s have a look at some historical Tech case studies and their conclusions.

Cisco, Meta and Apple: Three case studies for what happens when the story breaks

The next step is to ask: if some or all of that circular piece of revenue disappeared or was repriced, what would history suggest about the potential drawdown?

Three examples are useful: Cisco in the dot-com bust, Meta in 2022, and Apple in its 2015–2016 wobble.

Cisco: Vendor financing meets the end of the bubble

In the late 1990s, Cisco used Cisco Capital to finance telecom and internet companies that bought Cisco routers and switches.

By around 2000, roughly 10% of Cisco’s 20-billion-dollar revenue came from deals that relied on vendor loans and leases. As long as credit was plentiful, that helped drive very high growth and justified a premium multiple. When the dotcom and telecom bubbles burst, customers failed, funding dried up and capex collapsed.

Cisco took close to one billion dollars in provisions on bad loans in 2001 and wrote down billions of inventory. In one fiscal quarter, sales fell by around 30% year-on-year. The stock dropped roughly 85%–90% from its peak near 80 dollars to lows around 8–10 dollars. Cisco survived as a business, but the stock never really recovered the 2000 peak in real terms.

On one hand, Cisco’s case has many similarities with Nvidia’s one: a considerable part of revenues that is visibly circular, financing structures make demand look stronger and more durable.

On the other hand, there is an important difference regarding where the risk sits. With Cisco, the vendor loans were on balance sheet and produced direct credit losses. With Nvidia, the capital stack is more layered: Nvidia provides equity and preferred capacity, but banks and external investors provide most of the debt and incremental funding. That shifts more of the risk from “credit losses” to “revenue and valuation”.

Meta: A small revenue decline, a huge multiple reset

Meta in 2022 is the cleanest example of how little it takes on the top line to trigger a huge move in the stock when it happens at the same time as when a Tech hype ends.

Meta’s revenue in 2022 declined by roughly 1% year-on-year; its first ever annual decline. Profits fell much more sharply as it coincided with an increase in its Metaverse-linked capex (Reality Labs) and subsequent investor disappointment on the return on that investment. The stock fell around 75%–77% from peak to trough.

The lesson here is simple: it doesn’t take a 30% revenue collapse to get a 70%+ stock drawdown. It “simply “ takes the market to decide that previous growth isn’t coming back and that capital is being misallocated.

If investors start to believe that 10%–25% of Nvidia’s revenue was essentially a by-product of aggressive equity-fuelled expansion, and that this slice is now at risk, the Meta case study can become as relevant as the Cisco one.

Apple: When a revenue scare stays a scare

Apple in 2015–2016 provides a useful contrast.

In that year iPhone revenue declined for the first time in years; full-year revenue was down in the high single digits. The stock sold off by roughly 30% peak-to-trough. However, over time, the market re-rated Apple as it demonstrated that the ecosystem and services (Appstore commissions) growth were intact.

Putting it all together

For Nvidia, the Apple-style outcome maps to a world where the circular part of demand represents over-ordering and aggressive expansion that ultimately gets backfilled by “real” AI demand from enterprises and consumers. The Cisco/Meta outcome is the world where that circular part is revealed as structurally-overstated demand that doesn’t get backfilled on the same terms.

How big a market correction could an AI implosion trigger?

First, let’s define an “AI stock group” that consists of companies whose revenues are tied to a large extent to A growth:

• Hyperscalers and large AI infrastructure: Nvidia, Microsoft, Amazon, Alphabet, Broadcom, Meta, Tesla.

• AI-heavy Cloud and software: Oracle, Salesforce, ServiceNow, Palantir, Adobe, Snowflake, Datadog, Cloudflare, Workday

• AI semiconductors and equipment: AMD, Micron, Applied Materials, Lam Research, KLA, Qualcomm, ASML, Marvell, Intel, TSMC, Texas Instruments, Analog Devices, NXP

• Critical AI design and security names: Cadence, Synopsys, Palo Alto Networks, CrowdStrike, Fortinet, Zscaler

This group of stocks currently represents more than 30% of S&P 500 and more than 60% of Nasdaq-100.

In practice, AI is now the market.

Cisco / Meta Scenario: The bubble bursts and the narrative breaks

Assuming that Nvidia, and subsequently, the AI stock group fell by 75% on average (i.e. something like the Cisco or Meta peak-to-trough drawdown) and that the rest of the market remained flat (which is generous; in reality, it would also take a hit), then S&P 500 would drop by about 30% and Nasdaq-100 by 50%.

Apple Scenario: Growth scare, not full implosion

Assuming that Nvidia, and subsequently, the AI stock group fell by 30% on average (i.e. something similar to Apple’s drawdown around the iPhone scare) and that the rest of the market remained flat, then S&P 500 would drop by just over 10% and Nasdaq-100 by about 20%.That is a big correction, but not a crisis. It’s what you might see if the market simply decides that AI growth will be slower and lumpier than the current narrative suggests, but not that the whole thing is a bubble.

Wrap-Up

Nvidia’s circular deals with OpenAI, Anthropic and CoreWeave are large enough to matter. On plausible assumptions, they imply a Circular Amplification Multiple of roughly 2–3x for the model-lab deals (and far higher when you add leverage in cases like CoreWeave).

The Circular Revenue Ratio – the share of Nvidia’s data-center revenue tied to customers it also funds – probably sits somewhere in a 12–25% range. That is not a rounding error. It’s 2–3x the circular revenue share Cisco had at the dot-com peak.

The AI Stock group now accounts for roughly one-third of the S&P 500 and two-thirds of the Nasdaq-100. That concentration means that whatever happens to this circular slice of AI demand will propagate straight into the major indices.

The distribution of outcomes is wide. An Apple-style growth scare could mean a 10%–20% index hit. A Cisco/Meta-style unwind could mean something closer to 30%-50% drawdowns, depending on the index.

From an investor perspective, the question is not whether AI is “real” or “fake”. It clearly solves real problems and will be with us for decades. The question is how much of today’s infrastructure demand is 1) subsidised by equity and cheap credit, 2) pulled forward by aggressive cross-investments, and 3) priced as if current growth rates will persist long after the subsidy stack is gone.

The two metrics in this piece, CAM and CRR, give you a way to track that over time. As Nvidia announces new strategic investments, guarantees or capacity deals, they will need to get updated.

On top of that, there are a few events and signals worth watching:

• Changes in Nvidia’s disclosures around strategic investments and guarantees.

• Any hints on customer concentration by type of buyer (hyperscalers vs model labs vs neoclouds).

• Funding conditions for GPU clouds and AI-heavy data-center operators: spreads, deal volumes, willingness of banks to lend against GPUs.

• Evidence on AI unit economics at the application layer: are Cloud AI services and copilots generating healthy, sustainable margins or still being subsidised?
  

As the AI cycle matures, the market will eventually have to separate what was equity-boosted infrastructure build-out from what is durable, end-user-driven demand. When that happens, hopefully this analysis will have helped you come out stronger as an investor.

See you next week for more AI insights.

Share

Last week, Artificial Analysis released their quarterly AI report for Q3 2025 covering language models, image/ video models, speech/ audio models and hardware.

Have AI models reached a performance plateau? Are open-source models catching up with frontier models leading to commoditisation? Can frontier labs do equally well across text, video and speech? What gaps do hyperscaler and AI frontier labs have that can be addressed by AI startups?

Let’s dive in.

📃 Executive summary

Here is our summary of the report.

General comments

• Frontier labs continued to push AI model capabilities. The “intelligence frontier” remains fiercely competitive, with no clear dominant winner yet.

• Suggestions of progress stalling in AI are “greatly exaggerated”.

• Proprietary models continue to lead; open-source models are improving, but still lag at the very top.

📖 On language models

• The latest frontier model launch (GPT‑5) of OpenAI reclaimed the lead with an Intelligence Index score of 68. Artificial Investor note: expect Gemini 3 to take over the lead this week.

• Pricing for inference has been falling significantly (about 50% or more in many “intelligence classes”) as newer models and competition drive costs down.

📺 On image and video models

• Video generation models saw rapid quality improvement.

• Image editing models (especially instruction-based and multi-image input) are gaining adoption, e.g. Google’s Gemini 2.5 Flash.

• Chinese labs are moving ahead in video generation.

🔈 On speech/audio models

• The report introduces the new AA-WER (Word Error Rate) index for transcription accuracy measurement across difficult conditions - the lower the score, the better.

• The leading transcription model is Chirp 2 by Google, at 11.6% WER. The leading open-source models are in the 13-14% WER range.

• Native “speech to speech” models (audio in → audio out) are becoming viable and attractive for voice‐agents.

💻 On accelerators/hardware

• The hardware arms race continues: for instance, the new NVIDIA B200 chip in the report’s tests shows substantial performance gains (3x throughput vs the previous chip generation, H200) in their system load test.

• Inference demand has risen significantly: reasoning models, longer contexts, agent systems all drive compute demand up.

• Some challengers in hardware (e.g. Intel’s next generation chips) are delayed.

✍️ Reading between the lines

📈 AI infrastructure costs decline, but total AI costs increase

The page is making a simple but important point: we’ve made GPT-4-level intelligence 100x cheaper, but we’re also spending it even faster.

On the left, you see where the savings came from:

1. Smaller models and sparsity: Smarter training and better data mean we can get away with lighter models, giving roughly a 10x reduction in compute per query.

   • For instance, Mixture of Experts architectures like the one used by DeepSeek make it possible to have a large model with 700 billion parameters while only using 40 billion parameters at any time.

2. Software efficiency: Model optimisations squeeze another 3x efficiency.

   • For instance, the FlashAttention-2 paper validates a 2x - 4x speedup over baseline models.

3. Hardware efficiency: New accelerators that increase compute efficiency contribute another 3x cost improvement.

   • For instance, benchmarks see NVIDIA’s Hopper (H100) leading to 6x faster training of GPT-style transformers vs its previous generation chips (A100).

On the other hand, the demand side completely changes the picture. We’re not just running the same old GPT-4 queries more cheaply. We’re scaling up how much we ask the system to do:

1. Larger models push 5x more compute per query as we keep following scaling laws to get more capability.

   1. For instance, Meta’s Llama 3.1 model suite included a 70-billion-parameter model for “large-scale AI-native applications” and a 405-billion-parameter model for maximum performance. As compute per token scales linearly with parameter count, this implies a 5x compute demand increase.

2. Reasoning models “think out loud”, generating far more intermediate tokens, driving 10x more tokens per query.

   1. Most public examples of Chain of Thought vs “just give me the answer” show answer lengths going from a one-line response (i.e. 20-30 tokens) to multi-paragraph reasoning (i.e. 200-300 tokens).

3. AI agents don’t stop at one call; they chain many calls together to complete tasks, leading to 20x more requests per use.

Net-net: efficiency gains are real and huge, but applications are racing ahead, pulling total compute (and therefore infrastructure spend) sharply upward again.

🎥 Startups and scaleups are competing well in the audio/ video AI market

The audio/ video AI stack is highly fragmented: dozens of labs (from OpenAI and Google to Runway, Pika, Luma, Kling, Kuaishou, etc.) cover slightly different combos of text-to-video, image-to-video, editing and audio.

Elo charts (currently the most reliable performance benchmarks) for both text-to-video and image-to-video show that startup and open-source models (Runway Gen-3, Hunyuan, Wan, LTX, etc.) sit right alongside or just behind proprietary incumbents.

In other words, there is no “GPT-4 of video” yet - performance is clustered, competition is wide open, and specialised video labs are able to match or beat hyperscalers on many benchmarks.

The speech AI stack is also highly fragmented: a mix of big generalist labs (OpenAI, Google, Amazon, Meta, etc.) and a long tail of speech-focused players all offer overlapping speech-to-text, text-to-speech and even speech-to-speech models.

On accuracy, the Word Error Rate bars are tightly clustered: leading systems sit around 11–15% WER, with many others not far behind, and only a few laggards above 20%.

Net result: speech quality is already “good enough” across a wide range of vendors, with no single runaway winner and plenty of room for niche or specialised models to compete.

🐘 Google is picking up momentum

Survey data shows Google Gemini has gone from side character to co-lead: adoption/ consideration has jumped from 31% to 80%, putting it just behind OpenAI and well ahead of most others. At the same time, Meta Llama and Mistral have been sliding.

When you map the full value chain, Google is the only player with real strength in all four layers: apps, first-party foundation models, Cloud inference and its own accelerator hardware.

That end-to-end stack makes Google uniquely defensible: it can tune models to its infrastructure, push them through its Cloud and distribution, and capture value at every layer as Gemini demand compounds.

Plus, Gemini 3 is out this week. The jury is still out, but given that it’s released 4-5 months after OpenAI’s GPT-5, it should get Google back to the top of the model benchmark tables.

💭 The Artificial Investor’s thoughts

The fact that AI costs continue to rise despite all efficiency gains is really important for AI founders and investors. On the Infrastructure Layer side, this is good news as it indicates that as Intelligence penetration grows in businesses and in the daily lives of consumers, the demand for AI will continue to rise.

On the Application Layer, things are a bit more complicated. There are a number of founders that are running applications on very thin or negative gross margins to maximise their market penetration, whilst not being concerned about their business model: “AI infrastructure costs are declining and will continue to do so, so I am not concerned about gross margins”. While this is true, we are also witnessing overall AI costs increasing instead of declining. So, what is true?

Actually, both are true. Here is the catch: These businesses will see their costs decline only if they pause and stop innovating. So, if I were an AI application founder, I would ask myself: “can I win the market using the current generation of open-source models for the next 2-3 years? Or do I need better models, better reasoning and stronger agentic capabilities to make my product stronger?” If the answer is the former, then you are on a path of growing profitability and a sustainable business model. If the answer is the latter, then don’t expect your costs to decline, thus gross margins to improve, any time soon.

Also, the evolution of the video and speech AI model market is great proof that no company can do absolutely everything very well.

To go deep in a domain you need the right specialist technical talent, dedication of resources and management focus. The frontier labs are chasing AGI (Artificial General Intelligence) and the ingredients that seem to be capable of getting us there: language intelligence, logical reasoning, self-improvement and multi-step actions (agents).

The infrastructure and vertically-integrated segments of the “AGI” market currently seem “occupied” by the cash-rich leading players. However, there are other markets, as we are seeing in the case of audio, video and speech, that are still up for grabs. Examples include physical (robotics), Edge (machine-to-machine, sensors, etc) and numerical intelligence. This is where we see great opportunities for AI investors and founders.

See you next week for more AI insights.

Share

OpenAI’s CFO said that she would support federal measures to facilitate the financing of AI chip investments. This gets interpreted as a negative signal on the company’s financial health and helped trigger a 1-trillion-dollar sell-off by public investors. At the same time, the Chinese frontier lab, Moonshot AI, released its new open-source thinking model, Kimi K2 Thinking, claiming that it beats OpenAI’s GPT-5 and Anthropic’s Claude Sonnet 4.5 in benchmarks across agentic reasoning, search and coding.

Are we in an AI bubble? What would the impact be of a potential bubble burst? How significant is Kimi’s model launch for the industry? Is the US still ahead of China in the AI race? What does it mean for AI founders and investors?

Let’s dive in.

🗞️ Investors take 1 trillion dollars of value out of stock markets scared off by OpenAI CFO comments

📢 What’s the fuss all about?

In a Tech event organised by the Wall Street Journal, OpenAI’s CFO, Sarah Friar, said that she would support federal measures, such as a government backstop, to make it easier to finance investments in AI chips. Her quote was interpreted by some journalists and investors as if the company is near financial troubles, resulting in a broader AI stock sell-off. In total, 1 trillion dollars of value was taken out of stock markets last week, of which 800 billion dollars was related to Tech (thus AI) companies.

💭 The Artificial Investor’s thoughts

First of all, market volatility is the new normal and has been the new normal since 2021 for various reasons that have nothing to do with AI (such as inflation, interest rate fluctuations and geopolitical uncertainty). Specifically with regards to AI, this is not the first, nor the second…it’s actually the fifth “AI scare” since the “ChatGPT moment” in November 2022. We had two more AI scares this year - the “MIT 95% AI project failure scare” last July and the “Microsoft Capex scare” last February. Also, there were two AI scares in 2024 - the “DeepSeek scare” at the year end and the first “AI ROI scare” in July 2024. As with other AI scares, public markets reacted erratically and shaved hundreds of billions of dollars of value off Tech stocks within a few days. We do believe in AI’s long-term impact on productivity, but this doesn’t mean we are not in a bubble and or that the path is laid on roses.

We talked about the current hype around AI a couple of weeks ago, making references to discussions we had with Tech leaders and investors in Silicon Valley, but we didn’t talk about a bubble. Now is the time to talk about it and probably the best starting point is Azeem Azhar, the author of The Exponential Age, and his recently-introduced “AI Bubble Dashboard”. Azeem uses five gauges that capture different dimensions that indicate whether we are in an innovation-driven economic bubble or not:

1. Economic Strain: Is AI investment “sucking” the whole economy?
   The metric is US AI/datacentre capex as a percentage of US GDP. Today it’s about 1% of US GDP and green. It will turn amber if it climbs above 2%.

2. Industry Strain: Are revenues keeping up with capex?
   The metric is the multiple of AI/datacentre capex over GenAI revenues. Currently at 8.5x and in the red zone.

3. Revenue Momentum: How fast is the revenue growing?
   The metric is the number of years it takes to double GenAI revenues. Currently this is at 0.8 years, which is high, thus green.

4. Valuation Heat: Are prices detached from reality?
   The metric is Nasdaq’s valuations (tracking Price/Earnings multiples). Currently this is 35.2x vs. a normal figure of around 20x and the 72x reached in the dot-com bubble. This is now amber.

5. Funding Quality: Is the money patient and resilient?
   This metric is qualitative and has to do with the type of money that is pumped into the sector. Today’s buildout is largely backed by hyperscalers’ cash flows (good), but we recently saw Meta’s launch of an “AI bond” (private credit), which increases the risk. We are now at amber.

Azeem’s current conclusion is that we are not in an AI bubble, as we need at least three red gauges, but we currently have only one.

We believe things are a bit more negative than that, as the picture changes when overlaying a layer of profitability, and tracing government funding. The Industry Strain becomes 20x+ after applying an average gross margin of <40%. Revenue Momentum gets into the amber zone when converting it into Gross Profit Momentum, thus, taking into account the erosion of gross margins in the industry. Finally, hyperscaler funding is matched by US public funding (CHIPS Act, project Stargate, etc), which is linked to a government that is overlevered, with debt reaching 125% of GDP. We consider the Funding Quality gauge red. Overall, we like Azeem’s dimensions, but our interpretation is that we are much closer to an AI bubble, as we have 2 red and 2 amber gauges.

Finally, another valid question is what would happen if a bubble were to burst? The Economist published an interesting article about this topic making comparisons to the dot-com bubble focusing on the evolution of two key metrics:

1. The share of US household wealth in American stocks. This was 17% in 2000 and has increased to 21% in 2025.

2. The share of technology in the top 20 S&P 500 companies. This was 29% in 2000 and has increased to 52% in 2025.

So, we have an S&P 500 that would get impacted more by a sharp decline in Tech stocks and an American population that would get impacted more by a sharp decline in the S&P 500. This would indicate that there would be a tougher landing vs. 2000 if an AI bubble were to be created and burst.

The Economist concludes that a dot-com-style crash (a 75% decline from peak to trough) would cut American consumption by about 890 billion dollars, i.e. about 3% of GDP.

For reference, the trend is similar in Europe, but the figures are significantly lower:

3. The share of EU household wealth in European stocks. This was 8% in 2000 and has increased to 13% in 2024.

4. The share of technology in the top 20 European public stocks. This was 13% in 2000 and has increased to 16% in 2025.

🗞️ The new Chinese model Kimi K2 Thinking beats GPT-5 and Claude Sonnet 4.5 in benchmarks

📢 What’s the fuss all about?

Moonshot AI quietly posted a technical page for Kimi K2 Thinking, an open-source, agentic LLM built on a 1T-parameter MoE (≈32B activated) that’s trained to reason step-by-step and autonomously invoke tools.

On its model card, Moonshot reports:

1. About 45% on Humanity’s Last Exam (HLE), which is ahead of the American frontier closed models

2. The ability to perform agentic tasks consisting of 200-300 sequential steps calls

3. A context window of 256k tokens, double the one of OpenAI’s cutting-edge model, and

4. A version of the model that is twice as fast and more energy-efficient without losing precision

Naturally, the global AI industry was taken by surprise, sang the praises of the Chinese frontier lab and rang the alarm on the US-China AI race.

💭 The Artificial Investor’s thoughts

It’s hard not to feel the DeepSeek déjà vu here: once again, a Chinese open model lands with outsized performance-per-dollar and steals the market narrative. Remember when R1’s “20 - 50 times cheaper than OpenAI’s o1” claim helped trigger a sector wobble at Christmas 2024?

A few conclusions from our side:

First, this is another proof that “constraints lead to innovation”. Chinese model developers do not benefit from access to Nvidia’s super-performant AI chips. Instead of intimidating developers, this constraint has clearly bred a ruthlessly efficient design that enabled the K2 team to do more with fewer GPUs.

Second, we’re past the point of debating whether open models can beat closed on important areas: between Kimi K2 Thinking posting state-of-the-art agentic scores (HLE, BrowseComp, long-horizon tool use) and the broader shift, that we have witnessed as investors, of European developer mindshare toward Chinese open releases, the evidence is stacking up.

Third, this is further proof China is effectively on par with the US in critical AI dimensions: doing more with less compute, mobilising domestic ecosystems around chips, models and infrastructure, and shipping fast enough to force strategic recalibration in the West.

Finally, from a technical point of view, K2 is not a simple model improvement; rather, it’s the starting point of a new trend in model optimisation. Devansh at Artificial Intelligence Made Simple wrote an extensive piece on K2’s model architecture back in July 2025 and summarised it nicely recently: “Moonshot AI beat hyperscalers without the hyperscaler investments thanks to a trick called Muon.”

Elevating this at 30,000 feet, GenAI models consist of millions or billions of numbers called weights. Models get trained with a lot of data in order to find the best combination of weights, so that their result is close or even better to what a human would produce. An optimiser is the rulebook for how to change those numbers after each mistake. You can think of it like this:

1. The model makes a guess.

2. We see how wrong it was (the “loss”).

3. The optimiser says: “Okay, change these numbers a bit this way and that way so next time we’re less wrong.”

So, different optimisers = different styles of learning.

As per Devansh,”the choice of optimiser in deep learning is not trivial. The industry standard, AdamW, has been adopted widely because it is stable and low-risk, yet comes at the cost of efficiency. Kimi adopted a different optimiser, Muon, which is much more efficient but introduces intrinsic instability”. This is the reason that it has not been used by any American frontier lab.

Kimi’s innovation comes in the introduction of a variation of Muon that they called MuonClip, which overcomes Muon’s inherent instability. This is revolutionary in the field of deep learning and will likely be adopted by more AI developers in the future.

💵 What are the implications for Tech founders and investors?

This is great news for anyone building across the agentic AI stack, given that underlying models are improving at a fast pace. If the model’s stability over hundreds of tool calls proves true and becomes reproducible, then that’s a step toward trustworthy autonomous research and coding operations, not just one-shot prompting.

Also, this is good news for builders on the application layer that are betting on cost reduction to make their business model more viable. Implementing multi-provider brokering between K2 Thinking, DeepSeek, Meta’s Llama series, etc. will help optimise latency, cost and quality per step. The rise of provider listings (e.g. Together AI and OpenRouter) makes this commercially practical.

See you next week for more AI insights.

Share

We spent a week in Silicon Valley having 20+ meetings with leaders across the AI value chain (chips, datacentres, foundational models, Enterprise solutions, consumer hardware and consumer software apps) from virtually all hyperscalers, as well as with leading venture capital funds and Tech investment banks. This is the second part of our trip wrap-up.

Before we get to it, a few quick words on the main AI advancements of last week. The five US hyperscalers reported Q3 results. The common denominator for the three Cloud providers (Microsoft Azure, Google Cloud and Amazon AWS) is that they are all 1) growing fast (39%, 34% and 20% year-on-year respectively), 2) supply constrained - with energy and AI chips mentioned as the bottleneck - and 3) with the exception of Microsoft, advancing with their proprietary AI chip initiatives. Meta beat market forecasts and reported 50 billion dollars of quarterly revenues, but a one-off tax expense reduced its profitability which led the company stock to a massive 11% decline in one day. Apple had a better day as it reported record revenues for its Services business, driven primarily by App Store commissions. This is the power of a platform: it allows you to sit back and watch others fighting over AI users while you collect platform fees.

In other news, Anthropic committed to purchasing 1 million Google AI chips (TPUs) and 1 million Amazon AI chips (Trainium 2) throughout 2026, sending signals about the market’s shift towards AI chip fragmentation (which we also mention later in today’s issue). This didn’t put off Nvidia investors who pushed the stock up and made Nvidia the first company to be worth 5 trillion dollars. This came on the back of several announcements made by Jensen during the company’s annual GTC event in Washington, including that in 2025-26 it will see its first frontier AI chip manufactured end-to-end in US soil. The company also announced a 1-billion-dollar investment in Nokia to help the company with its AI connectivity businesses.

Now, back to our Silicon Valley trip and part 2 of the wrap-up.

Subscribe now

🧱 Infrastructure Layer

What is mentioned in the results announcements is also felt on the ground. The market currently is clearly supply-driven, meaning, if companies had more AI infrastructure, they would have more revenues. And while energy resources are often mentioned in the press, our sense was that at the moment the bottleneck is AI chips. And for now, everyone is prepared to purchase them at any price.

Another important theme in the Infrastructure Layer is that there is a trend towards a fragmented market.

First, as we have read in the past, every hyperscaler and large AI company is working either on its own AI chip or on partnerships with chip vendors other than the market leader, Nvidia. This makes sense for any large company that wants to control its own destiny and cannot allow itself to depend on a single vendor. Also, hyperscaler competition brings pricing pressure, which impacts margins. And we all know that Nvidia is making software-like gross margins (nearly 75%) despite being a (primarily) hardware company. A supplier’s gross margin is a customer’s opportunity (through vertical integration), which means Cloud providers are targeting Nvidia’s margins.

In any case, this transition will take a few years, because 1) it takes time to manufacture chips at scale and 2) custom chips (developed by Cloud providers) will be better suited for stable demand and predictable workloads, but currently the market is changing very fast.

Second, despite the fact that for now Cloud and datacentre providers are focused on meeting demand at any cost, at some point in the next 12-24 months we expect gross margins to become the name of the game. And when this happens, the focus will be on efficiency. There is already innovation happening across many dimensions that focuses on efficiency, such as model compression, more efficient chips, etc. Nevertheless, one fast and relatively-easy way to achieve huge efficiencies is to use 1) cheaper, entry-level GPUs and 2) cheaper CPUs for certain workloads. This points once again to a more fragmented AI chip market. By the way, moving from a training- to an inference-oriented world also means that cutting-edge expensive GPUs (that are mainly developed by Nvidia) are needed less.

Finally, many regulated enterprises are still operating on premise. AI is becoming a catalyst for their transition to their Cloud, which is bringing further demand for AI infrastructure.

🖥️ Application Layer

The consensus here is that the horizontal and basic vertical consumer/ prosumer market is pretty much taken up by hyperscalers (mostly Google) and large-scale AI companies (mostly OpenAI. Perplexity and Cursor). We believe this is not exactly true, as we have seen cases like Lovable, where an AI prototyping application was launched after ChatGPT, and, despite competition from Replit and Bolt initially, and now OpenAI and Google, reached 100 million euros of ARR within 18 months. Also, another case to consider is Superhuman, an email app (yes, an email app in a world dominated by Outlook and Gmail) that launched in 2018 and within 5 years reached 100 million dollars of ARR. Fyxer is another example of an email company that is growing fast towards €20m ARR within two years of its founding (October 2023).

Nevertheless, the bar is very high in terms of capital needs, product quality, speed of commercial execution, and community building. After all, Google is a great company in infrastructure, data and AI, but not as good at building consumer products that people love. As a result, most Silicon Valley founders are now building vertical applications (see also the latest Y Combinator batch) with Financial Services and Healthcare being the main verticals (is Fintech coming back?!).

Entry barriers are also a focus of AI builders and investors in the Valley. The typical software entry barrier of “being 12-18 months ahead of others starting now” is getting vaporised by the speed products can get built with the help of AI. So, what matters most now is speed of innovation, speed and quality of product iterations with clients, having a distribution edge and, of course, building a brand.

This discussion goes beyond defensibility (building a product that is difficult for others to copy) and reaches the point of independence: protecting yourself from the scenario of OpenAI and Anthropic tripling their prices tomorrow; a scenario that we see as plausible and analysed in last week’s issue. One way to achieve this is vertical integration, which can take different flavours: from training your own base large language model, to using proprietary data to either finetune a base large model or train from scratch a small model. We talked to some investors that would only back vertically-integrated applications.

By the way, on the AI agentic side, it came up in several conversations that currently businesses are buying like crazy “AI Augmentation”, i.e. agents that fit into existing workflows and help humans become more productive, but are not buying yet “Autonomous AI”. The reasons are 1) the technology’s precision is not there yet, 2) larger businesses need to rewire their processes and workflows to enable Autonomous AI, which takes time and effort, and 3) executives don’t fully trust AI yet.

🔮 Emerging Tech trends

We took the opportunity to ride a Waymo, the autonomous car of the ride-hailing company spun out of Google, and our experience was amazing. In some ways I realised that a ride from a hotel or my home to a meeting location is a very private thing and an autonomous car (despite the fact that I am sure it has a microphone and camera recording everything), felt much more private than a taxi or Uber. I could set the exact temperature and put the exact music I wanted (connecting my Spotify account) at the volume I wanted. When I had to make a confidential phone call it felt natural (I didn’t feel I had to use codenames). Also, as weird as this sounds: I liked the fact that I didn’t have to talk to anyone! I enjoy a lot talking to people, particularly in cities where I don’t live, to hear their story and get the latest local news, but when I am moving during a work day from one professional meeting to another, I mostly value privacy.

I also made a couple of other observations. First, AI is not driving in a perfect robotic way that we see in demos of video games: very patient, perfect speed, perfect turns, etc. It is driving like a human and, whilst it always obeys the rules (speed limits, stop signs, etc.), in more complex cases, such as changing lanes or navigating roundabouts, it behaves like a human driver. It’s perhaps hard to explain, but when you ride a Waymo, you will know what I mean. Second, it has an assertive driving style. Multiple times, I found myself on the right lane of a two-lane street behind a car that was going slow and the Waymo accelerated and overtook the slow car and went in front of it. I was not expecting that from a machine. There “was no reason” for doing this, as I would have arrived at my destination regardless. Perhaps it was a human teleoperator that took over the car from the AI; but if I was impressed.

Waymo is now operating in 7 major US cities and just announced they will start testing in London in the following months. Nevertheless, we are still a few years away from “autonomous cars happening”. If by “happening” we mean that the majority of cars are self-driving and we people stop getting driving licences, then we are many years away. From a technology perspective, the solution is not completely ready. First, the human-in-the-loop component is still critical, as human teleoperators intervene relatively frequently, second, the service does not cover the entire city of San Francisco for unknown reasons (Andrej Karpathy, ex-Tesla data scientist and early OpenAI employee) thinks it’s because there is bad network signal that prevents the teleoperators from connecting, and, third, it still doesn’t cover edge cases, such as someone placing an object on the car (some activists had placed cones on Waymos as a way to protest). Let’s not underestimate the time it takes to get from “works 90% of the time” to “ works 99% of the time”. Every “ nine” in the percentage accuracy of a tech solution can take even more effort than the previous “nine”. Beyond the technology limitations, we also have regulations, taxi driver unions/ lobbies and public opinion that will certainly play a role in slowing down adoption.

We also took the opportunity to try out the new Meta Rayban Display glasses. We were absolutely wowed! The first impressive functionality was the ability to talk to someone and then see real-time captions of what they say, translated to the language you want. The same-language captions are very fast and the translation feels fast enough to be practical in daily usage. This is due to AI models running on a local processor - Edge AI at its best! The second impressive functionality is the wristband. The glasses come with a neural wristband that picks up your hand movements and allows you to navigate menus, turn up and down the volume of the glasses’ audio and even generate text by handwriting on an imaginary paper. Apart from the convenience of making it faster to take some actions (vs. spelling them out), this functionality also addresses (to some extent) an important limitation of Voice AI: confidentiality. You could write a message to someone (without using a phone or laptop), and without people around you hearing the content of the message you sent. It’s fantastic.

There are some use cases that we can see having early-adopter demand, such as business meetings and any kind of travel that require frictionless translation, elderly who find it challenging to operate on a small mobile phone screen and people with visual or hearing impairments. At the same time, Amazon has identified a good use case for their delivery professionals that typically have their hands full and has recently announced their own version of smart glasses.

Clearly the current bottleneck is the battery technology. Battery life is 4 hours and the batteries are not swappable, which means you have to stop wearing the glasses to charge them. This compares to a smart watch that has a battery life of a day and takes a couple of hours to recharge.

From an aesthetic perspective, the glasses are a bit bulky. Not that the extent that we (and other early adopters/ geeks 🙂) would not wear them, but perhaps to the extent that style-sensitive critical mass would struggle. The reason is the size of all required components (battery, microphones, cameras, projector, etc), which are becoming smaller, but are still not at the required size.

As we wrote in previous issues, this is Meta’s real chance to dominate an end-to-end independent platform (as opposed to Facebook which is indeed a platform, but it sits on top of the Mobile and App Store platform, on which Apple makes high commissions). Nevertheless, it will still take some years until 1) the battery life gets to smartwatch levels, 2) the size of the glasses is exactly the same as the one of standard Raybans, and 3) the price of the “display + wristband” glasses goes down to the simple smart glass levels (200-300 dollars). Once these 3 things happen, we expect to have mass adoption.

Finally, we had conversations about quantum technology. How could it be that there is such a wide deviation in the responses to the question “how far are we from Quantum happening”? Some people responded 1-2 years whilst others 5-7 years away. The reality seems to be that both are correct. We are very close to the first use case at a small scale. But, when we talk about all banking passwords being decrypted by a quantum machine, then we are still some years away. By the way, Google announced just a few days later that their new quantum computer, Willow, solved a very specific math problem 13,000 times faster than a conventional super computer.

🐰🐢US vs Europe

Finally, one thing we noticed is that the gap between the US and Europe from a Tech sector perspective seems to be widening.

There is a compound effect of the combination of 1) 5x the capital injected, 2) large Tech companies breeding entrepreneurs that start new Tech companies, and 3) a large, unified and regulatory-light market that has been going on for at least 12-15 years (since the last two hyperscalers, Tesla, Meta, went public).

As a result, we now find ourselves with the US hyperscalers collectively investing more than 400 billion dollars annually on AI infrastructure, the US government announcing Stargate, a 500-billion-dollar AI infrastructure project and its AI Action Plan that involves supporting the country’s AI leadership with deregulation, Nvidia becoming the first company with 5 trillion dollars of market value, and Silicon Valley darlings, such as OpenAI and Anthropic, driving a large part of the AI innovation and reaching 10 billion dollars of revenues within 2-3 years of their commercial launch.

Where does this leave Europe? Europe has indeed been the birthplace of several big Tech outcomes, such as Spotify, Revolut, UiPath, etc. What gives Europe its edge? Where does it make sense for founders and investors to focus their time, energy and capital in this AI wave? We will cover this topic in a later post.

See you next week for more AI insights.

We spent last week in Silicon Valley having 20+ meetings with leaders across the AI value chain (chips, datacentres, foundational models, Enterprise solutions, consumer hardware and consumer software apps) from virtually all hyperscalers, as well as with leading venture capital funds and Tech investment banks. Given that last week’s highlights were anticipated product launches (OpenAI launched its AI browser, Atlas, while Google launched its “vibe-coding” app within the AI Studio), we thought it would be more valuable to share our learnings from our trip.

Let’s dive in.

🫧 AI hype is a fact

Clearly a recurring conversation is whether we are in an AI bubble or not. We won’t go into such conversation, as it’s quite complex and requires a thorough economic analysis (perhaps for a future post). However, we are comfortable talking about hype. And there is definitely an AI hype.

On the way from the airport to the city one can see dozens of billboards: AI SDRs, AI for Customer Service, Enterprise AI platform, AI models by the token, etc.

On the corporate side, we talked about::

Corporate leaders see AI as the way to cut costs and increase profitability. So, leaders mandate their managers to implement AI (no matter what). This means that most companies are trying out multiple AI applications on the same use case to learn and eventually choose one. The result is a surge of demand for Enterprise AI applications.

This is also the case on the consumer side, where the drivers are word-of-mouth, enthusiasm and fear (of staying behind), but the resulting surge in demand is the same.

As a result, companies with strong founders and engineering teams achieve unprecedented growth rates: 0 to 5 million annualised revenues in weeks or months. We discussed a lot about how permanent this surge in demand is. The consensus was that this demand peak could be temporary, but in any case it is a great opportunity to leverage the reduced sales cycles to gain market share fast, which, assuming you build a product that is sticky and with strong entry barriers, creates long-lasting value.

💰 How and when will AI make money?

At the same time, in conversations with company CFOs and Tech investment bankers we talked about whether the profitability of the traditional software model is at risk or not.

On top of the traditional software direct costs (cloud hosting and customer success costs), we find the AI infrastructure costs, which come in various forms: from model training and fine-tuning costs to paying “AI Model as a Service” fees. For now, end-user applications have not passed on these costs to their customers and, as a result, Application Layer gross margins are down to 50% in the best case scenario or 10% in the worst cases.

Source: Bessemer Venture Partners, The State of AI 2025

Another thing we talked about is how AI end products are not just underpriced; but also priced in a very unsophisticated way. For instance, within the 20 euro/month subscription I can ask ChatGPT about a recipe for tomorrow’s dinner at home, which has a value of 20 euro cents, or I can upload my blood test results, Fitbit data and eating habits of the last three years, and get a personalised nutrition and wellness advice that is worth 600 euros. Does it make sense?

So, will long-term growth rates be higher and thus offset the impact of lower margins on valuations? Or will we see lower exit valuations for AI companies in the future?

Correlation of valuations with revenue growth

Many Silicon Valley-based founders and investors are not concerned about gross margins at the moment. Perhaps it’s because of the historical and expected AI infrastructure cost trend.

In addition, in many conversations it came up how every hyperscaler has either developed their own proprietary model (Google Gemini, Meta Llama, Microsoft Phi, Amazon Nova, Apple Ajax, Nvidia Nemotron, IBM Granite, etc.), while working on their own AI chips (more on that in a later post). It is, therefore, reasonable to assume that AI models will get commoditised and, thus, so will AI infrastructure. As the costs of AI infrastructure decline, they could eventually get built into the Cloud costs resulting in software margins again. Great news for AI companies!

Not so fast. Anecdotes from various meetings indicate that AI corporate adoption is relatively slower than the pace of AI enthusiasm. The top 10% most advanced US enterprises are currently using AI at scale and get a positive measurable ROI out of it. We also see great productivity gains in our working environment. I personally see great productivity gains in my personal life. So, the value of AI is real. But if only the top 10% of US enterprises are at this stage, we are probably about 2 years away from mainstream business adoption in the US and 3-4 years in Europe.

Where does this leave us?

⏱️ The clock is ticking

Overall, it is as if we had two reverse clocks running at the same time:

1. The AI Subsidisation Clock: This is the hyperscaler and VC capital that subsidises the build-out of the AI infrastructure and thus guarantees demand

2. The AI Commoditisation Clock: This is the supply of cheaper AI chips and the launch of cheaper AI models that lead to long-term sustainable infrastructure prices

If the AI Subsidisation Clock stops before commoditisation takes place, then the infrastructure providers may demand higher prices from application providers, who would subsequently get squeezed between infrastructure and end users. This may lead to market collapse.

If the AI Subsidisation Clock runs long enough for the AI Commoditisation Clock to catch up, i.e. for sustainable prices to catch up, then we get wide AI adoption.

Source: The Artificial Investor

We have a couple of recent Tech innovation examples that went either way:

1. Solar energy: Subsidies and learning curves pushed capacity past scarcity into a commodity equilibrium: cheap panels, low margins and massive deployment.

2. Fiber and Internet boom: Internet infrastructure (fibre cables) was overbuilt ahead of demand, the prices of bandwidth collapsed and we had a wave of bankruptcies. Many years later fiber became a commodity input and the value shifted to apps and services, but a market collapse preceded that.

🏅 Early winners

So many unanswered questions to figure out the end state and winners and losers of the current AI trend.

For now, it appears that some winners have already emerged:

1. Some LPs (investors in VC funds) have “benefited” from this AI wave, as it helped them support their investment in VC funds (despite declining liquidity)

2. Some VCs have “benefited” from this AI wave, as it helped them raise their new funds (despite lack of returns)

3. Some founders have “benefited” from this AI wave as it helped them grow very fast and believe in big outcomes again

4. Some large Tech companies have “benefited” from this AI wave, as it helped them find ways to change their competitive dynamics or invest their large cash reserves in a strategic way

See you next week for the second part of our Silicon Valley trip wrap-up.

Subscribe now

A recent Yale study analysed the US labour market in the 33 months since ChatGPT’s launch and did not find a significant impact of AI, when compared to previous years. This is in contrast to other studies from Stanford and UK’s ISE.

Figure AI unveiled its first fully-integrated humanoid robot designed for mass manufacturing. Figure plans to begin commercial deployment in logistics and manufacturing environments next year. Meanwhile in China, Kepler Robotics announced that it has begun mass production of its humanoid robot, Kepler A40 Hybrid.

Another study on adults across 25 countries around the world found that one in every three are more concerned than excited about the increased use of AI in their daily life, while another 40% are equally concerned and excited.

What does this all mean for founders and investors in AI? What are the challenges and opportunities?

Let’s dive in.

🗞️ Yale research indicates that AI is not impacting the labour market

📢 What’s the fuss all about?

Yale’s Budget Lab ran an analysis on the US labour market in the 33 months since ChatGPT’s launch and concluded that AI is not having a noteworthy impact. The analysis shows that while the occupational mix is changing, and more so for younger employees, it is not a large difference and predates the widespread introduction of AI in the workforce. This is seen in the above chart that analyses the dissimilarity index for college graduates comparing 2015 (pre-COVID and pre-ChatGPT) and 2022 cohorts (on-ChatGPT). The study also found that AI exposure, automation and augmentation show no sign of being related to changes in employment or unemployment.

This is an important topic as it will set the tone on politicians’ stance towards AI, which affects public policy and pace of innovation.

💭 The Artificial Investor’s thoughts

This study comes in contrast to the Stanford study that analysed US payroll data and showed that graduates into white-collar jobs where AI penetration is high, such as software development and customer services, have suffered a headcount decline, while other age/experience groups saw an increase in the same period.

Similarly, a report from the Institute of Student Employers (ISE) had found that the UK Tech industry experienced a 46% drop in graduate jobs vs. 2024.

Our discussions with decision makers in SMEs indicate that AI is already impacting their appetite to hire graduates in internship and entry-level positions. Typically in these positions there is an unofficial trade between “teaching one how to do a job”, which costs the company, and the “low-level work” carried out by graduates. However, companies increasingly see that AI can do some of that low-level work faster and cheaper than humans, which brings them to the natural question: “why should they hire entry-level people then?”

On the other hand, our conversation with key AI decision makers and operators in large Enterprises indicate that they have made a conscious decision to continue hiring graduates, for now. However, the reason is not that the outcome of their own AI cost-benefit calculation is favourable for humans; rather, it’s because it is too early to make such an important decision. They prefer to see first 1) if their AI investment will continue at these (high) levels, and 2) the graduate skills gap will persist.

Ultimately, we believe AI will significantly impact jobs, but in a slower and more fragmented way than some pessimist journalists predict.

Companies operating in high-growth sectors are being pulled by market demand and tend to try to do more with the same resources (plus AI). Also, companies operating in sectors where personnel cost is a small fraction of their total costbase (e.g. ones treating with physical assets), are not focused on implementing AI. And then, even for companies with high personnel costs in a slow-growth sector, where profitability is a priority, their AI adoption will vary significantly with size. Large enterprises have legacy systems, complex organisational structures and processes, are subject typically to more regulatory oversight, etc – all factors that slow down technological adoption.

So, we do expect AI to transform the job market, but we don’t expect this transformation to be homogenous.

🗞️ Humanoid robots are moving from testing to scaling

📢 What’s the fuss all about?

Two major recent announcements show that humanoid robotics has moved from prototype to production.

In the US, Figure AI unveiled Figure 03, the first fully-integrated humanoid robot designed for mass manufacturing. Built on a joint AI system with OpenAI, Figure 03 combines visual perception, reasoning and dexterous manipulation in one end-to-end platform. The demo video showed the robot understanding spoken commands, identifying objects and completing complex tasks in real time. Figure plans to begin commercial deployment in logistics and manufacturing environments next year. For that purpose, earlier this year the company opened BotQ, a robotics manufacturing facility with initial capacity to produce 12,000 humanoid robots per year.

Meanwhile in China, Kepler Robotics announced that it has begun mass production of its Kepler A40 Hybrid, a humanoid robot using a combined electric-hydraulic architecture for strength and endurance. With a price tag around 25-30 thousand dollars, Kepler’s model costs a fraction of Western equivalents and represents one of the first large-scale humanoid robot manufacturing efforts globally.

💭 The Artificial Investor’s thoughts

This marks a turning point: after years of demos from Boston Dynamics, Tesla and Figure, humanoid robotics are moving to the industrial scaling phase. Figure’s partnership with OpenAI gives it an AI brain that can reason across vision, language and action. Kepler, on the other hand, demonstrates that China can scale robotics hardware faster and cheaper, leveraging its supply-chain dominance, component ecosystem and government support.

In terms of the US/China AI race, the parallels with the EV industry are striking: the US leads in model intelligence and software stack, while China leads in manufacturing speed, cost efficiency and vertical integration. If this pattern repeats, the first large-scale humanoid deployments could well happen in Chinese factories, not American warehouses. And, assuming Europe remains open to products from both countries, Chinese humanoid robots may likely have a higher market share in Europe than their American competitors, as is the case with new EV purchase currently.

Now that robotics are no longer a moonshot domain, what are the implications for AI founders and investors? What are the opportunities and challenges?

While most sector-agnostic horizontal manufacturing markets are already highly competitive (to not say closed), there are a number of white spaces where innovative companies, including European ones, can thrive:

1. Middleware and orchestration layers connecting LLMs with robot control systems

2. Simulation, data generation and synthetic training tools for robot learning

3. Robotic fleet management solutions that ensure coordination of robots deployed in the field and integration and cooperation with their environment

4. R&D in specialised actuators, batteries and materials optimised for humanoid endurance

5. Safety, compliance and auditability frameworks for human-robot collaboration

European companies’ proximity to Automotive, Aerospace and Industrial robotics clusters (e.g. Germany, Italy, France) may provide the right conditions to compete in industry-specific vertical automation.

🗞️ People are generally more concerned than excited about AI

📢 What’s the fuss all about?

Pew Research Center, a nonpartisan, nonadvocacy fact tank that conducts public opinion polling, carried out a study with adults across 25 countries that showed that overall a median of 34% of people are more concerned than excited and 42% are equally concerned and excited about the increased use of AI in their daily life. Older adults, women, people with less education and those who use the Internet less are correlated with higher concern levels. In countries such as the US, Italy, Australia, Brazil and Greece, the majority of adults are more concerned than excited about AI (comparted to the global medial of 34%).

💭 The Artificial Investor’s thoughts

Public sentiment in relation to AI is important, because societal friction can slow down adoption curves just as effectively as regulation. We’ve seen this before: in Social Media (privacy), Biotech (ethics) and Fintech (trust).

AI could be next. If half of the population feels uneasy about AI, this will shape regulation, corporate behaviour and, ultimately, where capital flows.

On the other hand, this may be a tailwind for startups that embed trust by design. Companies focusing on transparent AI, Explainability, privacy-first data models or auditable decision pipelines could gain a moat, in particular in high-stakes markets like Healthcare, Finance and Government, where “safe AI” is already becoming a key consideration.

See you next week for more AI insights.

Share

Link to the audio version of the post: