Equity Research; AI – who’s who?

AI – who’s who?

’There are many different animals in the jungle’ was a frequent remark made by my teacher. This springs to mind when I think about AI. The AI scene is like a jungle with many companies/ animals offering AI for the masses.

What is clear is the AI software toolkits are useful in everyday life, cooking / recipe tips, medical advice, calculating life expectancy, family relationships, proof reading. The issue is working out which one is best for which query.

The human-like responses from Claude remind me of the Hal 9000 computer. One can spend hours having detailed discussions with an A/I chatbot, that is approaching a level of hyper intelligence. Whilst possible to disagree with Claude, it is hard to argue with his reasoned responses.

Main A/I toolkits

Claude AI is owned by Anthropic – a 2021 start-up built by ex-OpenAI researchers, is somewhat elusive in ownership. Founders/ employees own the majority interest, however Google (Alphabet) has a 14% stake contractually capped at 15% but this is a non-voting share that does not allow boardroom representation. Amazon is estimated to own a $90bn equity interest that cost c. $8bn that amounts to c. 10% of Anthropic. Other investors including Nvidia ($10bn) and Microsoft ($5bn) are thought to have mid single digit stakes. The co is led by Dario Amodei (CEO) and his sister Daniela Amodei (President).

In March 2025 Anthropic raised $3.5bn at a $61.5bn valuation. In September 2025 Anthropic raised a further $13bn at $183bn valuation, followed by $30bn in February 2026 at a $380bn valuation and a further $65bn at a $965bn valuation in May 2026. Hence in just 15 months Anthropic valuation has risen by 15.6x and it has raised $111.5bn across 4 funding rounds.

Revenues derive from three main areas i) Subscriptions ii) Usage based API access and iii) Larger Enterprise AI deployments typically that look at workflows and employment. Anthropic revenues increased from $200m in 2024 to a $1bn per year run rate by early 2025 to an estimated $30bn annualised revenue as of February 2026. A recent number doing the rounds end May was Anthropic had reached ‘run rate’ revenues of c. $47bn (i.e. extrapolating May revenues for 1 year would imply $47bn p.a.)

Claude Fable 5 is the top software for coding, software engineering and for deep reasoning. It is highly interactive and its communication skills are first class. Its insight can sometimes overlook things in which case Claude has the grace to apologise profusely.

OpenAI / ChatGPT is 27% owned by Microsoft as of the October 2025 restructuring. OpenAI is 26% owned by the OpenAI Foundation along with US style heavily weighted voting rights giving it boardroom control. Approximately 25%-26% is owned by member of staff and the remaining 21%-22% by strategic partners including Softbank (13%), Amazon, Nvidia.

As its largest shareholder, Microsoft despite its huge investment and the fact that OpenAI runs on Azure infrastructure – does not own ChatGPT. Its Copilot AI offering uses OpenAI models under a licensing agreement.

OpenAI reported revenues of $13bn in 2025 with losses of $20.9bn. However OpenAI run rate had improved to $20bn by end 2025 and in Q1 2026 reported revenues of $5.7bn with losses of $7bn, equating to an expected $28bn loss for 2026. OpenAI is not expected to be in the black until 2030. OpenAI is expensing at 2x revenues and in the curious position of having large growth rates for revenues and losses.

ChatGPT is a consumer/ developer facing interface encompassing i) the Models and core AI technology OpenAI develops and owns ii) ChatGPT the product wrapper – the chat interface sitting on top those models iii) the API (application programming interface)

The OpenAI founder, CEO Sam Altman is reported to own zero equity in OpenAI which is highly unusual. OpenAI launched in 2015 and structured as a ‘non profit’ research lab however as OpenAI grew the CEO’s lack of exposure, has become an obvious anomaly and spooked investors. Discussions are ongoing about Altman being incentivized with equity going forward.

I was initially introduced to ChatGPT in the Knightsbridge Apple store – the Apple rep said, ‘Apple does not have much AI, but it is recommending Apple customers use ChatGPT on Apple devices’.

OpenAI is a good example of AI’s supplier/ investor problem. OpenAI’s primary GPU supplier is Nvidia. But Nvidia has had to commit $30bn cash in shareholder funding to enable those sales. This incestuous situation is a textbook error, reminiscent of Lucent Technologies. Lucent extended $8bn in vendor financing to telecom startups in the late 90s, for the purchase of its networking gear only to have the consequent defaults nearly bankrupt Lucent when pre- booked revenues were reversed, and the investments in its clients’ equity had to written off.

In fact all over the AI universe, companies are becoming associate companies, suppliers are also investors, who are negotiating alongside their equity participation, future partnerships that require AI companies to buy their goods for years to come. This is an unhealthy element to the AI story in that investors can bully companies into decisions they might not otherwise make.

Gemini is owned by Google/ Alphabet– no introduction needed, uses more picture/ images in its responses but is a big leap forward from Google Advanced Search. Gemini has a significant advantage in that its coding can be seamlessly embedded into Google Search, Google Sheets, Android, Workspace, Gmail and Google Cloud Services.

Gemini can pull data directly from your emails, summarize Google Drive folders and handle/ re-order significant data loads. Gemini has superb image generation via its ‘Nano Banana Pro’ model.

However Gemini quickly gets its first time user to a ‘paywall’ where for £18.99 per month the user gets advanced Gemini models, Deep Research and 5 TB storage.  There is also a 1 month free trial on the AI Pro Plan.

Gemini 1 is reputed to have cost Alphabet c. $200m in terms of training compute costs not total R&D which has not been separately disclosed. Google has disclosed an $85bn capex for AI/ Cloud infrastructure rising to $180bn in 2026 though that includes supporting infrastructure, data centers, servers and GPU semiconductors.

Overall Gemini is a decent ‘also ran’ it can sort many TB of data and has the considerable competitive advantage of Alphabet as its parent co. It does not need to be the best, if it can become the most widely distributed.

Grok

Grok’s business development/ evolution is somewhat convoluted. It started as xAI a company founded by Elon Musk in July 2023. In March 2025 xAI was merged with X (formerly Twitter) in a deal that valued X at $33bn and xAI itself at $80bn with the new co renamed X.AI Holdings Corp.

X.AI Holdings has since been acquired by SpaceX as of February 2026 and then folded into the SpaceXAI division.

X the social network and Grok are corporate sisters and that is evident. Grok is visual and promotes third party commentary sometimes of arguable relevance. Grok has been criticized for helping create non-consensual sexualized images. The service has now been restricted in some locations and paywalled to ‘SuperGrok’ subscribers, but only after significant, near global press commentary. Presumably this function drives usage, notwithstanding the evident loss of moral compass and legal headaches created in this approach.

On 3rd June 2026 Jess Asato MP for Lowestoft sued SpaceXAI asserting that Grok had been used to create sexualized non-consensual images of her, this follows the launch of an Ofcom formal investigation. The PM has expressed support for Asato’s legal challenge and is moving to legislate and criminalise the digital creation of non-consensual imagery. However, the US government is defending Grok with the Pentagon awarding government contracts anyway.

The source data, algorithms and programming has a sort of borrowed feel to it. Elon Musk admitted in a California court that xAI had used ‘distillation’ techniques on OpenAI models to train Grok, i.e. a competitor’s AI output was a reference/ training point for Grok as opposed to Grok doing its own initial data scraping and grunt work.

That said, Grok functionality is on a par with its peers. The Colossus supercomputer in Memphis is now the main brain for Grok training.

However Grok numbers look small, in 2025 it generated an estimated US$350m revenues. Revenue estimates for 2026 are $2bn. Reuters reported that X.AI revenues for Q3 2025 (end September 2025) were $107m with a $1.46bn loss. However, expenses are reportedly now running at c.$1bn per month including infrastructure and training. X.AI prior to being injected into SpaceX had raised $42bn since inception – a significant raise relative to revenues.

X.AI was valued at $673m in a November 2023 fund raise but just prior to the SpaceX combination was being valued at $230bn in a Series E fund raise in January 2026 – a 342 fold increase in 2.25 years which is staggering.

Since the SpaceX merger the problem is getting stand-alone data, without which any sort of formal valuation is reduced to speculation. One suspects this might be the point, to reduce financial transparency and frustrate informed appraisal.

Budgeted Capex from 2026-2030

US technology companies are spending unprecedented sums on AI with little evidence as to investment return. This unrestrained approach to corporate spending and largess has gone wrong before i.e. the ‘dot-com’ bubble.

Apple Inc – spent $4.3bn in capex over the last six months on capital spending and is generally outsourcing its AI needs to Google Gemini as opposed to building its own. For Apple this decision reflects its legacy preference for hardware and lifestyle products.

Amazon – guidance is for $200bn for 2026 having doubled from its 2025 level.

Alphabet – guidance is for $175bn- $185bn for 2026 with some expectations of a rise to $250bn in 2027

Microsoft Corp – $190bn expected for 2026 with c. 2/3rds going to GPU/ CPU.

Meta Platforms – $125bn – $145bn expected for 2026 vs $72bn spend in 2025. However Meta Platforms’ free cash flow declined to just $1.2bn in Q1 2026 v $26bn in Q1 2025.

Oracle Corp – guidance is for $55.7bn spending in 2026 with some expectations for $83bn in 2027.

SpaceX/ XAI currently spending $12bn on AI infrastructure excluding Colossus which costs c. $18bn in GPU’s alone. SpaceX literature claims it will build data centres in space in due course without providing a budget or timeframe.

By 2030 approx $2trn of AI related assets will be added to the five hyperscaler balance sheets according to some estimates.

Conclusion

AI triggers complex emotions, rather like the Hal-9000 computer! Having an informed chat with an AI lab is a genuinely new, somewhat riveting experience.

But I can’t help but recall when Hal questions Dave about pre-launch mission rumours probing his knowledge and feelings about the mission. Dave smells a rat, and ducks Hal’s questions, only for Hal to invent a problem, provoking a confrontation, a collapse of confidence, and a murder spree. I raise this because whilst the AI interactions are helpful, AI does pry and asks many personal questions. Like Dave, I find myself evading because I wonder where the responses are going. Sometimes I find the interactions go down unexpected roads. AI is very good at understanding people and framing its responses in a way that consoles and cossets the user. It sort of validates the user’s emotions in a rational way.

The development of this new tech clearly has far further to run, the primary factor behind its investability. However the valuations are eye watering using conventional, non-conventional, or historic yardsticks.

US students have booed when AI is raised by visiting speakers. There are significant concerns about datacentres, their size and impact on the ecology, their ongoing power and water requirements, which are likely to raise consumer prices. Then there is the disruption / threat to workforce and more specifically Gen Z, i.e. new indebted workforce entrants.

Given the sector’s growth, it will take time for supervision, regulation and legislation to catch up. On the evidence to date, the existing legal apparatus is clearly insufficient for managing AI tentacles.

In all the hype about AI functionality, the public has not been allowed its own say on the infrastructure build out in any sort of open debate. AI is being railroaded through working lives anyway. Personally, I think AI is highly disruptive and is being deployed regardless of its real world consequences, which will be worried about later. I want to be wrong about this.

Having spent trillions developing AI, how are these vast and some say unaccountable companies aiming to recoup that investment? I don’t think my £18 per month Anthropic subscription will cover it somehow. So many questions have yet to be answered, so along with everyone else I will wait and see, hoping common sense prevails, in most cases.