AI Rant: October 2026

Where are we today?

I have been building Decla for the past nine to ten months. In that time, I have seen AI progress from Opus 4.5 shipping a PR end to end in Claude Code to reports of 10,000 instances of an internal GPT model solving the Navier–Stokes problem. Any normal human looking at this progress would be—and should be—baffled. It took us “ex-apes” millions of years to develop intelligence, so watching this much progress unfold in months is hard to process. (No, I am not implying that these matrices are smarter than us.)

The volume of data these agents read and write is multiple orders of magnitude greater than what we are used to. I have personally had to delete my OpenCode sessions just to free up disk space on my 512 GB Mac. Frontier labs like Anthropic are moving away from tools like Git—designed for plebeians like us hoomans—toward tools like Scion (https://www.anthropic.com/engineering/redesigning-source-control), which can handle their shipping volume more efficiently. The game has changed, and software should now be built with agents in mind, too.

Everybody has something to worry about!

I saw this X post by Martin Casado of a16z about moats (https://x.com/martin_casado/status/2111851691524190279?s=20) and found it extremely relevant. Like some people on X, I think of the AI ecosystem as a stack of four layers. Each layer has its own advantages, and each faces new challenges as AI advances.

Layer 1: Energy

The first is the energy layer: Westinghouse and other players that provide the power and cooling solutions needed to run data centers. I see this layer as the most insulated from market fluctuations because you need energy to run anything. Period. Today, it is data centers; tomorrow, it could be flying cars!

Even so, having the most defensible moat does not make this layer immune to change. Large data centers could become obsolete if we discover an alternative to transformers that delivers comparable intelligence without the same need to scale. Scaling laws have been extremely kind to AI progress so far. These would-be Millennium Prize solvers are chunky, and the idea that bigger models become more intelligent has driven demand for compute—and, in turn, the energy needed to power those chips.

Still, these companies have been around for ages, and demand for energy extends well beyond AI. I think they are unlikely to disappear anytime soon.

Layer 2: Chips

The second layer consists of chip giants like NVIDIA, AMD, and Huawei, along with Google's TPU business. These are the players turning sand into the silicon that powers AI. NVIDIA started as a graphics processing company and pivoted toward AI as it became clear that deep learning math was extremely parallelizable. Chips like Blackwell, which introduced Tensor Memory—on-chip memory for storing intermediate tensors—and offerings like NVIDIA Dynamo and NVLink reflect that shift.

The natural question is: what happens to competition in this layer? Part of NVIDIA's success comes from the CUDA ecosystem and its early adoption by academia and industry. That gave NVIDIA a substantial head start, but I do not believe the advantage will remain as strong forever. No offense to CUDA or NVIDIA.

Other chip providers, especially AMD and Huawei, are catching up. Chinese open-source models like GLM are being trained on Huawei hardware, and LLMs are making it easier to write kernels. For instance, Sentryward (https://sentryward.io/) is a library that automates the formal verification of kernels, while companies like Sail Research (https://www.sailresearch.com/) are betting on delivering the fastest inference on any chip. Also, watch Neil's podcast!

So, if you are a newbie semiconductor company, should the existing competition scare you off? I would say no. Hardware and performance still matter, of course, but another company's lead in the tools used to program its silicon does not have to hold you back. This post gets at the same idea: https://x.com/MTSlive/status/2112215787608154267?s=20

Layer 3: Models and inference

The third layer is made up of model and inference providers. Here, we have giants like Anthropic, OpenAI, and DeepMind, alongside providers of open-source alternatives such as Z.ai, DeepSeek, and Moonshot AI. Competition in this layer is brutal. RL has been a boon, and every company is trying to outpace the others by producing the most intelligent model.

In 2024, people saw OpenAI as the leader with o1, which brought reasoning models into the spotlight. Then DeepSeek-R1 challenged that lead, and the team’s work on reinforcement learning for reasoning was later published in Nature (https://www.nature.com/articles/s41586-025-09422-z). The frontier then shifted to Anthropic with Opus 4.5 and its pairing with Claude Code, its native harness. Others soon followed, with GPT-5.6 giving the Opus models tough competition.

Around this time, Anthropic's Fable 5 (https://www.anthropic.com/news/fable-mythos-access) was temporarily banned by the US government. But competitors soon challenged Fable, including OpenAI's Astra models and Moonshot AI's Kimi K3. Even Anthropic's own Opus 5.5 is more intelligent than its Fable 5.1 offering (https://www.anthropic.com/claude-opus-5-5). Meanwhile, xAI's Grok models and recent Gemini models have caught up and are competing with the Sonnet and Terra models.

To me, this suggests that pure model intelligence is a difficult moat to sustain. I believe almost every model provider now has an offering that can handle most user workflows satisfactorily. Competition is increasingly about model costs, usage limits, inference speed, and the user experience—including personal assistants like Muse, Grokbot, and Dots.

For instance, Grokbot runs on a machine with Debian Linux 13, eight CPU cores, about 15 GB of RAM, and a 126 GB disk. Renting a comparable machine through AWS would cost around $260 a month, before model inference costs. Yet you can get Grokbot for $30 a month. I still do not understand the math behind that subsidized pricing.

Building a moat in this layer seems extremely hard, especially if you are a neolab competing with the incumbents. Hats off to you if you are building one. I have a lot to learn from you. (https://x.com/deedydas/status/2103903382822089010)

Layer 4: Applications

Then comes the application layer, aka the “GPT wrapper.” This layer gets a ton of criticism for its perceived lack of innovation. The argument is that one of the bigger labs could eat these companies at any moment. That may be true for some, but I do not think it tells the whole story.

For instance, Decla integrates with agency management systems (AMS), the systems of record agencies use to run their operations. One of them, Momentum AMP, recently introduced API pricing (https://api.momentumamp.com/). As more AI agents handle an agency's day-to-day operations, they generate more API traffic. Why not make some extra cash from that? Similarly, established businesses with valuable data could make it available through APIs and charge customers for access.

But what about AI B2B SaaS companies themselves? I believe there is hope here, too. If model intelligence is becoming a harder moat to sustain, application companies can benefit by building custom harnesses that work across models and perform extremely well in a specific niche. Say, insurance :D.

This Databricks blog shows how coding-agent performance varies with the model, harness, and effort level (https://www.databricks.com/blog/benchmarking-coding-agents-databricks-multi-million-line-codebase). I think the same idea extends to domains beyond coding. In fact, domains like insurance can be harder to optimize for because they lack the same established benchmarks. The designer has to decide what to measure and optimize.

An AI B2B SaaS company can still be competitive if it builds a harness that completes a workflow quickly and reliably while using less reasoning—and therefore reducing inference costs.

There is another opportunity in the application layer: building software that AI agents will use. Anthropic built Scion, an in-house version control system, but plenty of companies lack the expertise or manpower to build tools like that themselves. They may be happy to pay 20 bucks a month for tools their agents can use, much as they pay for services like Resend.

So, what should you do in the application layer?

If you are a legacy big dawg sitting on valuable data, make it accessible to agents and charge a fee. If you are building agents for specialized tasks, optimize your harness until you can deliver the same workflow at a lower cost than a $20 ChatGPT subscription would offer. And if you are building software, design for a future in which most of your customers may be agents!

Every layer has advantages to build on and new risks to confront. As AI adoption grows, each will have to tackle a new set of problems.

And what about the AI doomsday predictions? Is it over? I do not think so.

Have expectations increased—from VCs all the way to your manager? Absolutely. Just because a machine can do some things faster and better does not mean you are done. It means it is time to take on more challenging problems!