Infra Play #124: Best of 2025

The pivotal moment for cloud infrastructure software

It’s been quite the year. It’s easy to forget that last December, LLM reasoning was barely just launched, enterprises were still playing around with Llama 3.1 open-source deployments, and most software companies were still mostly talking about AI, rather than actually reshaping their software architecture around it.

More importantly, coding with AI was still seen as a weird thing that didn’t make a lot of sense, or at best was a slightly more interactive version of Stack Overflow.

Over the last 12 months, things have shifted dramatically, making the industry almost unrecognizable.

Microsoft and Google had a massive pivot in different directions, with the Azure team having an absolutely insane year and being on pace to catch up with AWS, while DeepMind delivered the best base multimodal frontier model on the market, trained on their own hardware. While both of these were not shocking developments, the fact that the teams behind them were able to execute at the level that they did has been nothing short of impressive.

Most of the large SaaS incumbents completely fumbled their AI opportunity, mostly due to a lack of vision, execution quality, and understanding of the right adoption mechanics. They looked at AI and asked “how can we run this at 80% margin”, rather than “how can we dominate the token usage across our audience”.

Speaking of dominating token usage, holy schmoly, what a year for Anthropic. After a difficult 2024, where their newly hired sales team barely got to $1B ARR after a significant push from the hyperscalers, Claude Code, Sonnet 4.1, and Opus 4.5 became the strongest product-market fit in Enterprise AI, rivaling the consumer hold that ChatGPT has. Projected to finish at $9B ARR, Claude has firmly established itself as the most widely used LLM for coding, and its most recent release Opus 4.5 is seen as operating as a competent (if occasionally on the junior side) developer.

At the end of last year, I wrote:

For tech sales: OpenAI’s latest mode o3 reasoning l is benchmarking at what we would consider AGI level for all practical purposes (i.e., it can replace an average worker). At launch, the cost per task is too high for mass adoption, but this is likely to change significantly within 12 to 24 months. If you are not currently working for a company well-positioned to perform in the new paradigm (i.e., a cloud infrastructure software vendor), you should seriously reconsider your options.

For investors: The majority of value always accrues at the bottom of the stack. The foundation of the AI stack consists of OpenAI, Anthropic, Meta, xAI, Databricks, NVIDIA, AWS, Azure, and GCP. The beta will reside at the top of the stack, with preferred data platforms, observability, and cybersecurity vendors that integrate within the hyperscaler ecosystem. This thesis required “conviction” in 2024; in 2025, it will become “obvious.”

…

Time in the market is significantly more important right now than timing the market. If your company isn’t doing anything meaningful with AI and lacks the vision + technical capability to progress quickly in that direction, you’re very exposed to disruption. The inverse, of course, also holds true.

If you’re in the right place today, you will be the disruptor.

This played out accurately in terms of value accrual and positioning. The definition of AGI shifted in the minds of early adopters as many realized that scaling the models both in terms of intelligence and production availability would require the largest infrastructure investment since, well, ever. At the time when this was written, an unreleased version of o3 essentially beat the ARC-1 benchmark, and it took a full year until we saw an actual production release from OpenAI that would match that score (albeit at a dramatically lower cost).

Speaking of OpenAI, there were very few stories over the year that weren’t connected to them in one way or another. There was a period where one announcement of working with the frontier lab was sufficient to add tens of billions in stock value by the end of the day. They finished the year at $19B ARR, becoming one of the largest and most successful cloud infrastructure companies in history, yet the revenue is seen as insignificant relative to the multiple commitments made over the year to acquire compute worth $1.4T. In order to pay for this, the company would have to take in over the next five years revenue above that of what we today consider “hyperscalers”.

2025 cannot really be covered properly without understanding the significant importance of deploying capital. What many categorize as the “AI bubble” was in practice one of the most concentrated shifts of capital in the shape of infrastructure buildouts and early-stage company investing. “The revolution will be televised” became a prophecy, as even a dedicated show on this dynamic called TBPN emerged.

Seed stage and Series A rounds exploded in valuations, as both VCs and founders tried to tackle some of the biggest challenges in the application layer. The lion’s share of capital, however, concentrated around the frontier labs and the infrastructure buildout that serves their needs. Companies like Oracle fully pivoted into building the nuts and bolts of deploying infrastructure, taking significant CAPEX risk in order to win the opportunity.

While AI became one of the most dominant day-to-day topics everywhere, there remains a massive informational asymmetry, which many are trying to monetize.

On one side is the “San Francisco consensus”, essentially the capital of forward-looking AI knowledge and access to the latest models. Adjacent to it is “AI Twitter”, the default social media of anybody who is willing to talk publicly about what they know and what they see coming. Slightly further away is Substack and the podcasting network on YouTube, the long-form content that often reflects either the “San Francisco consensus” conversations or the latest drama on the X timeline.

Everything outside of that sphere is behind. LinkedIn and corporate media are behind 3 to 6 months. The general tech public is behind 12 months. Everybody else is still living in a bubble where ChatGPT 4o and Microsoft Copilot are cutting edge and AI is just a fad.

The big challenge for anybody working in cloud infrastructure software is understanding where their network sits. The reality is that most companies are still led by individuals living at best 12 months behind. This lack of vision and capabilities translates into their poor roadmaps, floppy headcount decisions, and overall bearish stance on AI.

The next few months will be the last train to adjust for most. As Blackwell infrastructure goes live in the datacenters, model availability and performance will jump forward faster than most expect. Even those with access to significant capital like Meta are struggling to recover from poor decisions earlier in their strategy. What do you think will happen to your average SaaS incumbent who is still trying to figure out customer support chatbots that don’t suck?

Over 2025, I covered different angles of the industry, sometimes through the lens of its leading companies, sometimes through the mental models of its thought leaders. These are the best articles that I think capture the essence of how things played out:

The start of the year was all about NVIDIA and its thesis around the "AI factories for tokens" that will power the scaling of reasoning models across the world. What at the time was seen as "shilling from Jensen" proved exceptionally on point, as reasoning tokens exploded in usage.

My deep dive into Docker introduced the new vision of the company and warned of the perils of building GTM teams for critical, but difficult to monetize, cloud infrastructure software technologies.

Thinking about AI only in the context of the last few years is a very limited view. We are in a pivotal technology cycle, but it builds upon a number of other critical trends. Understanding whether this is a short-term expansion or a complete shift in the technology landscape is critical to hedging your bets.

Many of the players investing heavily in AI are not simply interested in ROI. There is a deep philosophical undercurrent behind some of those choices, and e/acc can be seen as the manifesto for what AI progress can look like.

At a time when the market was still bearish on Google, the big question was whether the current strategy would actually lead to a material change of perception, i.e., stonk goes up. In this article, I explored the thesis that ended up playing out in the second half of the year.

Continuing my exploration of developer tools over the year, my GitHub deep dive explained the rise and "fall" into enterprise adoption for the most widely used platform today to collaborate on code development. The article also includes one of my favourite founder quotes on building software that matters:

We cared about developers. But it wasn’t about when they added Git, it never really mattered.

They never had any taste. They never cared about the developer workflow. They could have added Git at any time and I think they all still would have lost.

You can try to explain it by the features or “value adds”, but the core takeaway that is still relevant to starting a startup today is more fundamental than if we had an activity feed or profile page or whatever. The much simpler, much more fundamentally interesting thing that I think showed in everything that we did was that we built for ourselves. We had taste. We cared about the experience.

We were developers and we built what we wanted in order to enable how we wanted to ideally work. We were the only tool in the space built by developers for developers without PMs or accountants or CEOs trying to optimize for revenue rather than for developer experience.

In the end we won because the open source community started to converge on distributed version control and we were the only ones in the hosting space that truly cared about how developers worked at all. The only ones who questioned it, approached it from first principles, tried to make it better holistically rather than just throwing more features onto something existing in order to sell it.

This is why GitHub won.

The launch of GPT-5 was a mixed bag in what was clearly a very successful, yet very difficult year for OpenAI on all fronts. The biggest shadow over the company was the negotiation with Microsoft over a new agreement between the two companies, which was resolved later in the year with a new deal. With this problem now resolved, the frontier lab could get back to the small challenge of defeating a resurgent Google, outperforming Anthropic, and Musk pushing aggressively to be the first to build the strongest short term infrastructure base in the business.

While the short-term focus on enterprise adoption of AI is what most companies and industry insiders care about, the "San Francisco consensus" is focused on different topics, including the most likely game theory of what happens in a scenario of superintelligence. One researcher decided to monetize this situational awareness and founded one of the most successful hedge funds of the year.

Oracle is a peculiar company because the public perception of it is tied to what is essentially legacy software for laggards. In the shadow of this brand, the company ended up building one of the most capable teams for deploying cloud infrastructure software, and in the last year emerged as one of the preferred partners of OpenAI for helping them scale datacenter capacity. The $1T valuation that followed sent ripples across the industry.

"Never bet continuously against geniuses" has always been a strong strategy to hedge your bets. Musk has defied all odds in pretty much every venture that he has been involved in. xAI, the frontier lab + social media company, is no exception. The trick is that building frontier models is a sport of kings, and at a time when both Anthropic and OpenAI showed massive growth, he has very little margin of error left from a financial perspective. This is where the importance of building the right go-to-market team becomes obvious, and where, I'm afraid, they are making a fumble.

While AI can often appear to be a lot of talk around benchmarks and coding tools, the political dynamics in the background are difficult to ignore. One of the interesting things about technology is that when a new shift occurs, its regional implementation might lead to significantly different outcomes. Understanding the Chinese point of view is critical.

Once most understood that AI was going to change how we use computing, the race to build applications started. Most ended up in "POC hell", not because the models were not capable, but due to how poorly the majority of enterprises actually handle data pipelines. If most enterprise data platforms tried to obscure complexity, Databricks made the contrarian bet of providing a complex set of tools that solved hard problems. $4.8B ARR and 55% growth later, the value of this strategy is becoming clear.

One of the ways to understand the dramatic shift that AI is driving across the industry is to look at what adjacent technologies and projects are being funded. Building datacenters in space is at first glance a bit of a meme, but it's worth exploring further.

Copilot is a struggling product; Windows became so unusable that a new wave of Linux adoption is happening across developers; Xbox is in a slow-motion exit from the console market. At a time when the public perception of Microsoft has never been worse, Azure is the undisputed top performer in enterprise AI+Cloud adoption. A lot of this success has to do with Satya Nadella and how he intends to play the next five years.

Cybersecurity continues to be one of the most important pillars of cloud infrastructure software. Funnily enough, the majority of cybersecurity companies operate under the same playbook of solving one niche problem with barely usable software, trying to monetize it, selling the company, rinse and repeat. In the shadow of causing one of the worst internet outages in history, George Kurtz has been reshaping CrowdStrike for the agentic era.

The turnaround of Google this year was driven by significantly improved execution in terms of building a frontier model and then productizing it. Behind the scenes, this accomplishment was underpinned by the contrarian bet of building TPUs, the custom hardware powering all of Google’s inference and model training infrastructure.

If 2025 was "dynamic", it's difficult to overstate how quickly things will move in 2026 if we see further jumps in performance from the frontier labs. As the scaling laws proved to be firmly in place with the launch of Gemini 3 Pro, the obvious need to see much stronger base models from OpenAI and xAI will depend on their ability to train them on NVIDIA Blackwell infrastructure. As the last 12 months were mostly spent deploying this advanced hardware and starting to train models on it, the real fruits of this shift will become obvious in the year to come.

This concludes 2025 from my perspective. Launching a cloud infrastructure software company database, writing 70+ articles and an ungodly amount of tweets have been time-consuming, but worth it in many ways.

We might not know what the ultimate outcome of AI will be, but maybe the real value is the friends we made along the way. See you next year, tech anon.