_{Disclaimer: All Moatt Black research and content is human written and produced. AI is only used for minor copy checks.}

How compute economics, throttling and variable intelligence could reshape the AI business model.

Social media, forums and platforms are seeing more and more users question if AI models are getting worse. Feelings of answers being more manic, more unhelpful and most importantly pushing users to conclusions faster are on the rise. At Moatt, we decided to investigate these reports and find out the reasons behind it, why this could be happening and produce our report on what this could mean for the future of AI, and ultimately the future of work.

Increasingly AI intelligence is not a fixed product but one that is a resource that is dependent on how the provider allocates compute, reasoning effort and latency, not only the underlying model that was created.

We feel this has major consequences for enterprise adoption, retail growth, and ultimately valuation.

The AI Director of AMD, Stella Laurenzo published a report in April 2026 based on a large amount of data from AMD’s internal engineering. The quantitative analysis covered 17,871 thinking blocks and 234,760 tool calls across 6,852 Claude Code session files.

Her conclusion on the state of Anthropic’s newest model was scathing:

Laurenzo’s extensive research published on Github was comprehensive in findings of Claude’s regressive behavior. With the findings broken down to four overarching factors:

Anthropic responded to this research with an investigation of their own, producing this postmortem in late April. The findings concluded three primary reasons for the degradation:

However, as this focuses around coding functions of Anthropic’s model, why are Laurenzo’s findings so important for the average user?

In a study by OpenAI on how humans interact with AI, the company found that a majority of 77% of messages fall into three categories: practical guidance, seeking information, and writing. By contrast, computer programming was only about 4.2%

Users are claiming that interacting with models seems significantly different, and difficult to get previous work completed with the same quality. On the 14th of August a Y Combinator Hacker News thread titled “Why does Opus 5 feel worse to work with?” drew 873 comments.

There is real data and evidence on opinion that AI models are degrading. So what is really happening, and how does it impact the future of AI, business use, and investment thesis?

The evidence itself does not prove a covert throttling of consumer intelligence. It does, however, demonstrate something economically more important: reasoning effort is an allocatable resource.

It’s clear now that intelligence can be delivered to retail and business users on a throttle. Large AI companies can and scale how much ‘intelligence’ they’re providing the customer at any given time, impacted by factors such as compute cost, speed of response and public opinion. In reason 1 of Anthropic’s postmortem of Laurenzo’s research we see a public version of a throttling decision, where Anthropic decided to change Claude Code's default reasoning effort from high to medium.

Google has also been public about its Gemini enterprise agent platform’s usage. Publishing an explanation on provisioned throughput. Additionally, in a study by Elioth Sanabria of Columbia University, entitled The Shadow Price of Intelligence, Sanabria discusses at length the possibility of throttling, and just how it can negatively impact the very issue it may be trying to solve.

The argument is by throttling output either in surge activity periods, or truncating content to limit compute costs often leads to users retrying queries increasing the problem that the throttling is trying to solve.

In March 2026 Anthropic and the US Government clashed over its model use being used to surveille domestic citizens. OpenAI ultimately stepped in and took up the US contract. At this point it seemed like Anthropic’s product Claude was going to gain massive market share. Many users changed from OpenAI to Anthropic some stating political alignment, and the majority claimed it was a better model, that it significantly improved work flow over ChatGPT.

In June 2026 you would have been right to have the feeling that the landscape felt like it had changed that OpenAI was fast becoming the next Yahoo, and Claude the next Meta. The current situation for Claude as it approaches a potential IPO is a difficult one, users are reverting back to OpenAI’s product ChatGPT and the good favor that Anthropic carried in Q2 of 2026 is rapidly being lost in Q3 and Q4.

The release and recall of Fable in June 2026 certainly hindered optics of Anthropic’s models too. Although mitigating factors from outside geo political impact cannot be attributed to the companies leadership. Anthropic handled this crisis communications event well, however much it may have further negatively impacted optics around the models.

Financial technology company Ramp recently analysed token data of Anthropic and OpenAI enterprise token usage and found that OpenAI is regaining enterprise spend against Anthropic, with Q3 showing enterprise token spend at OpenAI to be up 82% and Anthropic to be up 76%.

In research carried out by The Pew Research Center the study found that 35% of all content on the internet has signs of AI authorship since the release of ChatGPT in 2023. Additionally leading SEO firm ahrefs conducted a study of 900,000 new webpages and found that 74% have AI authorship.

The theory that we are seeing a degradation of data as AI trains itself on more and more AI written work is entited “Model Collapse”. There is real data to back up the theory, with examples from researchers proving that models forget the true underlying data distribution

In a research paper by Ilia Shumailov and Zakhar Shumaylov for Nature the pair state claims that LLMs are vunerable to model collapse in the way the models train and crawl new data.

It has been no commercial secret that pricing for consumer AI models are an arbitrary choice, chosen on emotional customer purchasing not hard data of costs and profit margins. OpenAI’s CEO Sam Altman has publicly admitted that pricing models were based on consumer emotional tests:

Altman has conveyed that back in 2025, OpenAI’s $200 Pro plan was losing the company money. Posting on X that “we are currently losing money on OpenAI pro subscriptions. People use it much more than we expected.”

Whilst there has been no data publicly released since this statement on the profitability of the Pro plan, changes to add an additional $100 pro plan and changed usage limits may signal that this is still the case.

This data showcases a similar and relative relationship to the early days of the gig economy, most notably with Uber.

Uber notably held a strategy to gain market share at any cost. The ride share platform lost $2 Billion a quarter to maximise the growth and market saturation it needed to fend off competitors in a race to its 74% market share, most notably with its main competitor as Lyft at 24% market share.

But what we’ve seen from gig economy platforms in recent years is whilst the company itself has become very profitable with Uber’s net income at $10 Billion for 2025, the industries and customers that support it are struggling significantly. A study by UC Berkley found California delivery drivers netted only $5.93/hour before tips. A study by Lendingtree found that customers ordering delivery are paying 80% more than picking it up themselves. A study by the National Restaurant Association found that full service restaurant profit margin sits at 2.8%.

We expect a similar dynamic to emerge across AI. As this research demonstrates, token limits and variable reasoning allocation introduce a clear mechanism for providers to manage compute intensity and unit economics. As AI becomes increasingly embedded in enterprise workflows, customer spend could expand materially beyond today’s subscription levels. At the same time, rising compute requirements will increase demand for the physical infrastructure that supports it, including energy, data centres, semiconductors and the critical minerals required to manufacture them.

The subsidy phase sets an expectation before the underlying economics have been fully tested. When OpenAI and Anthropic enter the public markets, the pressure to convert growth into durable margins will intensify. That transition will begin to reveal the true cost of delivering AI at scale, and by then millions of consumers and businesses will already be deeply embedded within their chosen ecosystems.

In order of appearance: