TLDR Highlights
The Shift
The "Honeymoon" is officially over. Enterprise FOMO has been replaced by scrutiny on ROI, and the "magic" of 2023 has faded into disillusionment.
The Threat
OpenAI is being strangled from both sides. Google dominates infrastructure (50% cheaper costs) and distribution (the Apple deal), while Anthropic has stolen 54% of the developer market.
The Tech Gap
The "God Model" era is dead. Google Gemini 3 leads in science/multimodality , Claude Opus leads in coding , leaving GPT-5.1 squeezed in the "good enough" middle.
The Money Pit
With a $14B annual burn rate and $250B in debt obligations to Microsoft, OpenAI is pivoting to risky hardware and ads to survive, potentially alienating users.
The Verdict
This is OpenAI's "Netscape Moment." Squeezed by incumbents and debt, the company likely faces a forced restructuring or breakup by 2027
As we celebrated the end of 2025, I recall thinking to myself how different the world was compared to just three years ago. I remember the first time a friend of mine told me about ChatGPT and the things it could do (he also said we were officially doomed).
Having returned to the UK from living in Shanghai in 2022, with memories of lockdowns and systematic temperature checks just behind me, I hoped the world would get back to normal. Boy, was I wrong.
The excitement in my friend’s voice when he explained exactly what an LLM was and what it did was almost infectious. I recall having a similar hope for my Commodore 64 when I unwrapped it and powered it on one Christmas morning. A younger me (I had an imagination back then) marvelled at the idea of a computer being alive and able to answer my questions. In hindsight, it was almost certainly a good thing that young me never got hold of AI.
However, as I learned more about ChatGPT, I felt a reminder of that younger version of myself as my imagination fired into life once again. I spent a lot of time thinking about my own limitations and how this could unlock new possibilities for me and my career. A small part of me even thought that life was going to get easier for a lot of people now and this might just be the future finally arriving.
Well, the future arrived. I paid for a subscription and away I went, asking it stupid questions, getting it to write short papers for me on nonsense subjects, etc. After a while, I began to realise there were limitations to this new technology and the magic began to fade slightly, not totally, but enough.
The more I used the tool, the more I grew frustrated with its quirks, and eventually, it became just another cool tool.
As we entered January 2026, we all got back to work and continued with our daily lives, a little older and wiser. The thing is, as each year has passed, the AI models have grown, sure, but they haven’t grown in parallel with our collective sense of novelty wearing off. They still do amazing things; the problem is, it’s not that amazing anymore, and the pace of innovation appears to be slowing.
The world is no longer as eager to drink the Kool-Aid as it was back in 2024, and it’s starting to show. The OpenAI bubble maxed out a while ago; we just haven’t realised it yet.
Let me explain.
The Macro Environment
OpenAI now faces a triad of challenges that have replaced the optimism of 2023-24.
In general, the Fear of Missing Out (FOMO) has dissipated almost entirely (if we exclude out-of-touch CEOs). Many corporations are seeing their much-hyped projects fail to reach production, creating “zombie pilots”. Increasingly, organisations are scrutinising spend on APIs that once promised the world. Productivity gains have been witnessed, albeit on an incremental basis rather than a transformative one.
Compute demand is strong, but physical capacity is unable to keep up. Unlike Google, which owns its stack (TPUs + Data Centres), OpenAI is merely a tenant in Microsoft’s Azure ecosystem. This makes it vulnerable to supply constraints and vendor margins.
The third challenge comes from regulatory bodies enforcing strict compliance. Enterprises are increasingly wary of “black box” models. Many companies are moving towards “stateless” solutions (where the vendor doesn’t retain the data), which undermines the training strategy of models like ChatGPT.
Competitive Threats
Google is the single biggest threat for OpenAI, and it’s only getting stronger.
They have the capacity to train and run models at roughly 50% of the cost of OpenAI simply because they own the hardware involved. Their TPUs (Tensor Processing Units) are made in-house and provide a cost basis that Sam Altman’s team, reliant on NVIDIA margins, cannot match.
Another catastrophic coup was the recent news of Apple signing up to power their dim-witted (I hate you, Siri) assistant, Apple Intelligence. The main reason reported behind the decision was Google’s willingness to run Gemini 3 in Apple’s Private Cloud Compute, a stateless environment. OpenAI, on the other hand, insisted on data retention, which was a deal-breaker that lost them potential access to 2 billion devices.
Furthermore, Gemini is allowed to train on YouTube videos, whereas ChatGPT is legally blocked from this vast data resource. When we see Gemini continuously creeping up behind ChatGPT, it looks increasingly like Google’s long-game strategy is paying off.
Let’s not leave out polite little Claude, either. This quiet assassin has now officially overtaken OpenAI in the mindset of software engineers. Latest estimates place Claude at 54% of the enterprise coding market, whilst ChatGPT sits at only 21%. The reason for the loss of this key user base is simple: accuracy and reliability. The future is not bright for OpenAI if coders build the majority of their apps using Anthropic’s API and not OpenAI’s.
The Old Benchmark King is Falling
The era of a single “God” model is ending; the landscape is becoming increasingly specialised.
- Google Gemini 3 Pro: The “Multimodal Dominator.” It leads in scientific reasoning (91.9% on GPQA) and video understanding. It offers a 2-million token context window, allowing companies to “dump” entire databases into the prompt.
- Anthropic Claude Opus 4.5: The “Coder’s Choice.” It holds the highest score (80.9%) on the SWE-bench Verified coding benchmark.
- OpenAI GPT-5.1: Squeezed in the middle. While it remains the gold standard for conversational fluidity and “agentic” tool use, it trails in deep scientific reasoning and long-context handling.
I wrote about this a while ago: today, choosing the right AI model depends entirely on your use case, and “one size fits all” is long gone.
OpenAI’s “Hail Mary”
To survive the commoditisation of its software and the loss of Apple’s distribution, OpenAI is pivoting to hardware and media.
Jony Ive is reportedly working on a screenless device with a voice-first interface that will act as a companion for the user. The hope is that this will help OpenAI break the smartphone chokehold that Apple and Google currently enjoy.
The main problem is that I just don’t think it will. To execute a global product launch with no existing supply chain network is hopeful at best. The other main obstacle is that we have seen voice-only AI devices attempted before, and they fell flat. Users are not going to change a long-standing method of device interaction overnight. Remember when Microsoft was convinced that users would love the touch interface of Windows 8? Didn’t that go well?
Despite Jony Ive’s talent, Sam Altman is no Steve Jobs, and he simply doesn’t have the endless product genius that Steve possessed. The device might sell well initially, as such things often do these days, but I do not expect it to change the game.
Therefore, Sam and his team must find a way to turn the assets they do have into a winning commercial success. With a roughly $14 billion burn rate, OpenAI is turning to ads to fund their enterprise. I cannot state strongly enough what an obvious own goal this is. It will likely accelerate their market share losses.
User trust is the single most valuable thing that holds people to their beloved ChatGPT. Once that relationship goes from trusted friend to used car salesman, the date is over. Google will be waiting with open arms to accept these fleeing users into its ecosystem.
Future Scenarios
OpenAI currently has $250 billion in debt obligations to Microsoft’s Azure platform. In the deal between OpenAI and Microsoft, if OpenAI achieves AGI (Artificial General Intelligence), the exclusive license is broken. This creates a perverse incentive for OpenAI to declare AGI as soon as possible, whilst giving Microsoft clear incentives to delay that event to maintain dependence.
Conversely, Softbank has invested a cool $40 billion in OpenAI, betting squarely on “Superintelligence” being achieved. That kind of investment always comes with pressure to deliver “moonshots”. Unless their bet on ads and the Jony Ive device miraculously pays off (it won’t), the future looks bleak for ChatGPT surviving unscathed in its current form.
In my humble opinion, we are about to see a Netscape moment for the modern generation. I fully expect the squeeze from Apple and Google, combined with the growing debt to Microsoft, to create a perfect storm. Within the next 2-3 years, I anticipate investors and debtors will force a restructuring of some sort, likely carving up the company.
The end result? OpenAI could well retreat into declining margins and user share, whilst Google and Claude carve out their solidified areas of dominance. Will Apple then make a move for Anthropic? That remains to be seen.
I wouldn’t bet against it, just like I wouldn’t bet on OpenAI surviving this arms race.