Artificial Intelligence (AI) promises could be dreadful or exhilarating
Introducing radical change for years to come with access to the largest body of knowledge conceivable, AI might be extraordinary
If the revolutionary transformation comes true, it is not just a massive wealth redistribution but a new social rapport which will emerge from the collapse of the old
Is society really on the cusp of radical breakthrough ? and are revolutions already breathing down our necks ?
Or could AI just be a 'marketing spin' to draw huge investment commitments, ending in a wimper of technological evolution, slowly digested by society?
I am not sure anyone has a definitive answer - the technology is truly spectacular but it is quite uncertain how the social forces governing our communal lives will respond
35 years ago, in his novel 'Jurassic Park',, Michael Crichton warned how the dynamics of scientific research did not guarantee diffusion of scientific advance for the common good
“You can do [research] very young. You can make progress very fast.
You don't even-know exactly what you have done, but already you have reported it; patented it, and sold it
As science gains in power, it proves itself incapable of handling the power. Because things are going very fast now..
What should I do with my power? - which is the very question science says it cannot answer."
Michael Crichton, lending his thoughts to a dying mathematician Ian Malcolm - Jurassic Park 1980
Precisely right - we have advanced science, we have powerful investors, but do we have a plan about what should be done ?
Valuations of all things AI are off the charts, for all we know
The supply chain is actually made of many things, and of many business models, from the shovels (and their wooden handles) to the 'gold' drawn out painstakingly by the prospectors
By considering those links one by one, some insights might be gained about where artificial intelligence (AI) stands as a business proposition for its investors
Science has been pushing the boundaries of what was deemed possible
The technology is already very consequential and no one can tell how far AI will run, transforming the way companies operate and individuals relate to digital reality
The Internet of the 1990’s might give proper insights about the dynamics of AI technological breakthroughs but no one really knows…
In this brief for the investor unfamiliar with the technological requirements but curious about investment opportunities, I frame my presentation in four sections
- The AI value chain and the business propositions of each link
- Have AI prophets gone ballistic, overselling the risk and the potential?
- The strategic importance of the ‘application layer’
- Are strategic early investors rotating their capital allocation within the value chain
The AI value chain
Artificial Intelligence is a value chain of single interconnected loops
The links are governed by business models operating under very different capital requirements and profitability expectations
Compute
Compute is the foundational base layer
Compute encompasses the physical hardware and processors, the specialized cloud infrastructure (data centers) required to train models and run AI applications, and the electric power
It is a capital-heavy infrastructure business, with usage-based server hosting, profitable for who can commit the capital expenditure
Roughly 70–80% of every dollar spent on AI inference is split between silicon providers (40-45%) and hyperscale data centers (30-35%) such as Amazon, Microsoft and Google Cloud
- The physical silicon chips that execute parallel calculations: Graphics processing (GPUs), Tensor Processing (TPUs) and High-Bandwidth Memory (HBM) are dominated by NVIDIA, Samsung, SK Hynix and Micron and foundries, with TSMC and more recently Intel
- In the AI semiconductor market, training systems require increasing amounts of high-bandwidth memory, while inference systems are increasingly optimized to reduce hardware requirements per workload, a moving balance of sorts...
- Networking and power generated by Utilities require hard infrastructures, drawing an estimated 10-15% of every dollar spent
The Model Layer
Model builders OpenAI and Anthropic are frontrunners in the West as generative AI providers, but Google DeepMind (Gemini) and Microsoft-AI also integrate AI natively in workflows (consumer or business)
Revenue generation has been massive, generated by token fees, enterprise licensing and consumer subscriptions, but GAAP losses derived from enormous 'compute' costs have been just as massive (recent declarations of Anthropic's 'profitability' notwithstanding)
Open weight, but not open source, AI models like Meta's Lama, X.ai and Mistral AI compete in the same space
In short, profits are to be found in the revenues posted by Compute, generated by demand of the Model layer
This is not about to change because every token consult generates a cost, which is the pole opposite of Internet search (generating no additinal cost at each consult, just profit)
It is unclear whether the model builders are in a race with just one winner or with a phalanx of specialized models
An additional factor to be considered is the emergence of various Chineses models, open weight and open source, with highly optimized training and smaller models, requiring less compute
Out of every dollar spent on AI, an estimated 10-15% is recorded in the gross margin of model builders
Competition in the 'Model Layer' and heavy financial losses are hard to square with current IPO annoucements (announced for the near future by Anthropic, delayed by OpenAI)
The Application Layer
Aiming at current dominance of SaaS (Software as a Service) firms such as SalesForce, applications are bulld on top of the AI models and have been marketing themselves as model-agnostic
The field is brutally competitive and even the curently most deeply integrated applications in the workflow of their client base may be 'squeezed' between frontier upstarts and model builders' applications
The most usual SaaS licensing model is priced and accessed on a per-seat (per user, per month) basis, which proves costly for the client
Instead of relying on external 'software services', the largest firms are already incentivized to migrate on internally developed AI applications, with the support of their own IT specialists
Smaller firms might follow, slowly crumbling the SaaS client base without any hope of recovery
What is left out of every dollar spent on AI may seem modest (3-5%) but profitability of the application layer is estimated at 70-85%
AI prophets gone ballistic
By a sleight of hands, Artificial General Intelligence (AGI) promises to be unique in the ethereal realm of ‘super intelligence’
The model builders, and foremost the IPO contenders, OpenAI and Anthropic, walk a tight rope, vaunting 'superhuman' performance while holding back (a bit) as self-promoting responsible actors of society
To captivate investor attention, the process looks familiar
- Just like during the Internet bubble of distant memory (in the 1990's...), AI models are believed to be on the cusp of miraculous profitability, with 'super-intelligence' and 'consciousness', neither of which is actually defined
- Applied to enterprise focus with Anthropic and dominance in AI chatbot universe with OpenAI, AI might drive exponential growth (and investor expectations) in adoption rates and in revenue
Predicted disruption sets a pattern where only company metrics matter now, and GAAP profitability will be achieved by dominance in the AI space later
In the latest rounds, investor finance has remained supportive
- $65 billion in May 2026 for Anthropic pushes valuation to $965 billion
- $110 billion in February 2026 for OpenAI achieving a valuation of $850 billion
Fighting over the application layer
To get 'real', the model builders have been making their business case with specialized agents as alternatives to the software services
Deployment by the models of AI 'agents' (for specialized tasks) and AI 'connectors' (to outside data bases) allow businesses to finetune software and replace standard CRMs (customer relation management) and even ERP (back-office Enterprise Resource Planning)
The legacy SaaS and analytics specialists have been repriced to reflect the disruption
- Market values of Thomson Reuters
, SAP , Salesforce , ServiceNow and many more firms focused on data analysis have fallen by the tens of billions USD - Productivity gains by companies deploying internal software stacks are probable
- Will such gains make palatable the costly subscription rates to the models providing the underlying infrastructure ? or does it remain a leap of faith ?
There is a scenario where the market cap losses of the SasS segment - exceeding a trillion dollars - are recovered without a hitch by the models, justifying their prospective valuation
Maybe... a doubt is creeping in this smooth scenario however, as agent-building suites are attacked from all sides
- Competition between model builders will be fought in a brutal game for survival
- Chinese open-source models are likely to undercut the low to mid-range AI software developers; Microsoft is testing a partnership with China's DeepSeek in this setting
- The most powerful independent AI applications will not tolerate anything other than a neutral model builder (capturing profitability for themselves)
-
The embattled SaaS providers advocate a hybrid approach, integrating their own AI frameworks onto their mature SaaS platforms as systems of record
Palantir
The most spectacular - and vocal - proponent of independent AI applications is Palantir, a case study in the reach of data integration and artificial intelligence infrastructure platforms
Palantir focused from the start on the entities, the relationships and the constraints describing the framework within which agentic systems can make decisions
Under the premise that this framework defines the 'reality' of what is 'allowed' and what is 'not allowed' in an 'ontology' layer, the company has created what looks like an unassailable moat, with secure digital systems consolidating massive, fragmented datasets for governments, militaries, and major corporations
By digging a trench with his usual strong-voiced argumentation, Palantir's founder Karp pairs its safe "ontology" application layer with open-source models (rather than Anthropic or OpenAI)
With a focus on 'trust', Mr. Karp raises the stakes for the generative AI model builders with a single question
"Who controls the data stack and to what purpose?"
Condemning AI model builders on principle, Mr. Karp spells out the consequences of their control over the data
Who controls the weights and the data stack, controls the means of production
Caching, temporarily storing data or model reasoning, will be ineffective in dynamic AI agent workflows because the caching mechanisms rely on exact text matches and cannot keep up. By running iterative loops and evaluating redundant tool calls from scratch, the agents will inflate costs by draining tokens
With data learning by the AI agents, transfer of proprietary corporate secrets and replication are a risk
The opening skirmishes in what could become a long battle between model builders and application layers confirm how fluid valuations remain at this stage
Nothing is cast in stone, not yet
Not for Palantir - with a trailing twelve-month (TTM) Price-to-Earnings (P/E) ratio hovering between 134 and 161, well above industry leaders such as Nvidia (PE 41) or Broadcom (PE 79) - topped by just one stock Tesla with a ratio fluctuating between approximately 327 and 359 (!)
Nor for the model builders eagerly awaited trillion-dollar IPOs
Rotating capital allocations within the AI value chain
Outsiders have the dubious privilege of attempting to catch straws in the wind
Having a data center in your backyard was a no-no 10 years agor (noisy, water-consuming, energy-hogging, small workforce and ugly)
For a short while, data centers have become a 'must have', everything an insightful, forward looking, tech spirited politician wanted in the backyard (of his constituency)
The weathervane seems to be pointing back again - and resistance of local constituencies is resurgent
Some of the largest data center announcements, Blackstone stands out
- terminating a planned Digital Gateway data center project in Virginia (800 acres)
- selling a 64% stake in 3 North Viriginia data centers, 100% leased to investment-grade hyperscale tenants
It is not as financial transaction but by valuation of data centers that Blackstone's sale stands out
- With rapid depreciation of 100 000's GPUs, the value of a data center rests on lease contractual coverage, air-tight quality of the lease and quality of the leasing counterparty
- The financing of greenfield data centers at the cheapest rates will be defined by the lease contract (take-or-pay and termination payment in case of early exit)
Another insight might be suggested by Brookfield, a major investor in U.S. data centers, with the firm's shifting attention to European data centers
The data center counterparties, committed to the multi-billion lease contracts, are the hyperscalers - Amazon, Microsoft and Google
Doubts may be of the tiniest sort, creeping in nonetheless...
Softbank, invested in OpenAI (a 13% stake, committing $64.6 billion), has reopened talks on $10 billion OpenAI margin loan, after being rebuffed
According to Reuters (July 2, 2026), SoftBank now offering corporate guarantee if OpenAI collateral proves insufficient
- Lenders in the consortium include Goldman Sachs, JPMorgan and Mizuho, sources say
All of this might be justified by rotation of capital allocations, but confident valuations of model builders and data centers are hard to find
In my next note, I hope to project AI valuation
Is it the swirling, contracting mass of a dark vortex...or just the opposite, an expanding universe full of promise ?