The market moved faster than even the most bullish expectations after the ChatGPT moment. Frontier models leapt from delicate demos to production systems. Open source models have become substitutes for enterprise workloads. Inference emerged as the dominant market in AI.

Underpinning all of this, AI compresses time. New models are released every 41 days. Companies reach $100m in revenue in record time. We all achieve more faster.

In celebration of our anniversary, we wanted to trace that mechanism through the market shifts of the last three years.

The first casualty of compressed time is the old language of venture capital. Seed, Series A, Series B categories still exist, but they describe the financial product companies seek rather than rather than company maturity. Venture firms have left the idea of offering a standard financial product to bespoke offerings : seeds range from $1m to $500m in size. Can we really call it all the same thing, anymore?

Three years ago, a seed company was often a small team with a product concept & early signs of product-market fit. Today, some seed rounds are larger than IPOs, fueled by great ambition, a supportive VC ecosystem, & the promise of generational scale businesses to be built.

Part of this is inflation in private markets. But more of it is time compression : the best companies mature much earlier than software companies did in prior generations. We’ve learned as an ecosystem how to build software companies & AI accelerates product development.

Compressed time also redraws the map of where great opportunity lies. When we first launched Theory, most AI conversations centered on models. Remember the debate of whether model companies would be the airlines of the era?

Today, inference is becoming the dominant market. The market is segmenting because the workloads & buyer preferences have evolved - very few companies can afford state-of-the-art AI for everyone - & each specialized constraint creates a new infrastructure category.

Companies like Sail Research are building the systems that operationalize intelligence : serving it cheaply, routing it intelligently, & specializing it around use cases like video, batch, local, agentic, & real-time workloads.

Databases followed this path a decade ago. They fragmented into OLTP, OLAP, vector databases, & streaming systems. Those markets have evolved with AI, a pattern we’ve backed through MotherDuck & LanceDB, with Omni in the analytics layer above them. Inference infrastructure is now specializing the same way.

The expense of inference reinvigorates a sedate market that has been controlled by behemoths for a decade : advertising. Every major interface shift, TV, web, mobile, streaming, found its answer to monetizing a massive audience in ads, & AI is no different.

AI advertising is emerging as the subsidy for inference costs, letting applications grow usage & revenue together rather than against each other. We wrote about this dynamic when we led Koah’s Series A : native ad formats inside AI conversations are producing click-through rates 4-5x the display baseline, & an agentic app builder can provide inference offset by ads.

The same compression closed the gap between closed & open models, cloud models & local models. The conventional narrative holds that frontier closed-source models lead & open source follows. We’ve reached the iPhone 15 moment of AI. Many models are good enough for most work.

Running a model locally reduces cost, improves latency, increases control, & minimizes data governance concerns. Enterprises are adopting local & open-source models for sensitive workloads, & frontier capabilities compress toward consumer hardware within a few years. What once required a hyperscaler cluster runs on a laptop just a few quarters later, a shift Ollama brings to millions of developers.

The promise of AI is that software will ultimately be more secure : machines that read every line of code, patch faster than attackers move, & never tire.

In the meantime, the attack surface is exploding. MCP servers, skills, plug-ins, & coding agents each introduce new entry points, & enterprises are deploying them faster than security teams can review them. Attackers are massively parallel & shrinking necessary response times from months to minutes. Defenses must respond.

It’s why we backed Dropzone, whose AI analysts investigate the alert flood no human SOC can keep up with, Maze, which applies agents to cloud vulnerability triage, & Artemis, securing the new agentic surface itself.

The same agentic wave is rewriting operations. ERP & back-office systems have resisted change for decades because the work is unglamorous, the data is messy, & the switching costs are enormous. One CFO we interviewed, when asked about a startup said, “that company has only been around 15 years; they are too immature.”

Agents invert that math. Systems that read documents, reconcile records, & execute workflows can attack operations from the inside rather than demanding a rip-&-replace. It’s the thesis behind Doss, rebuilding ERP for teams that move at modern speed, & Backops, applying agents to the back-office work no one wants to do by hand.

AI has impacted crypto, another market fueled by data. Prediction markets, stablecoins, micropayments all have an AI infusion to them. Today, crypto companies need to generate revenue & use AI to provide better experiences, which led to our investment Allium, the data layer underneath that institutional wave.

Recognizing shifts early requires fingers on keyboards, wrestling AI agents into compliance rather than observing it. We built Theory as a technical organization, experimenting with AI across research, sourcing, diligence, portfolio support, & internal operations. Working inside these systems sharpens our understanding of where the stack is breaking & where new workflows are emerging, while deepening our empathy for founders deploying real AI systems inside enterprises. It’s harder than social media says.

AI also changes the economics of an investment firm. Over the last decade, venture firms scaled by adding people. AI-native companies are demonstrating that much smaller teams can operate at 10x+ the leverage of prior software generations, & the same dynamic applies to us : since launch, we’ve analyzed 2x the investment opportunities with a team of just 3 investors working alongside a nine-person intelligence organization.

None of this works without the team behind it. Theory started three years ago as a handful of people & a thesis. Today we are thirteen strong. We believe this is the structure of a modern venture capital firm : engineers & researchers who build the systems we use every day : agents that map markets, pipelines that surface companies months before they raise, & research infrastructure that lets a small team cover the ground of a firm several times our size.

Everyone at Theory works with the technology we invest in, & that shared fluency shapes every decision we make. The firm we’ve built over three years is itself a product of the thesis : a small team, deeply technical, operating with the leverage AI makes possible.

But the real story of these three years is the founders. They compressed decades of company-building into quarters & shipped products that rewrote what enterprises expect from software.

The next three years will make these look slow. The most ambitious builders we meet are just getting started, & we can’t wait to see what they do.

I noticed the little llama icon on the door & remembered how much I loved the product. Michael opened the door & welcomed me in. That is how I became friends with the Ollama founders, Jeff Morgan & Michael Chiang.

Since then, I have migrated most of my workloads to Ollama ; it turns out I’m not alone.

More than 8.9m developers use Ollama, a number growing by nearly one million every week.

These users run Claude Code, Codex, OpenClaw, Hermes, & 65,000 more applications & integrations on top of Ollama.

Ollama makes AI simple. Download the app & start running AI on your laptop in seconds. Here it is running on my Macbook :

I run open source models locally for 70-80% of my work & then burst to the Ollama cloud for complex tasks ; a cloud that is twice as fast as others.

85% of the Fortune 500 uses Ollama across healthcare, finance, & energy. We found Ollama monitoring electricity loads at a powerplant in Finland ; at a space agency in the USA ; in the CFO’s office at a public company validating financials ; & even at a particle accelerator.

Data never leaves the user’s machine when running locally. Businesses can customize their AI & own it.

Ollama has partnered with every major open-source model provider including Google, Meta, NVIDIA, Microsoft, Zhipu, DeepSeek. Ollama offers thousands of models like GLM-5.2, Kimi K2.7, Gemma, Qwen, & Ornith.

Today we announced our investment leading Ollama’s $65m Series B, alongside Benchmark, YC, & others.

The future of AI is open models running everywhere work gets done. Jeff & Michael built Docker’s developer experience. Now they’re building it for AI.

If you want to run an open model, start here.

Context size is not the ceiling. Memory architecture is.

I have been experimenting with a memory architecture that runs preflight instructions. A pilot plans the route before takeoff. My agent does the same.

A query lands. “Summarize the Q3 board deck.” 200,000 raw tokens of emails, PDFs, & chats sit behind that sentence.

Preflight is retrieval. The agent inspects its skills library¹, picks the ones relevant to the task, & loads only those into the context window. Skills are consolidated memory ; the preflight step is how the agent picks the right one.

The local Ornith 35B model² then executes on that loaded context. Hard tasks route out to the frontier ; routine tasks remain on the local model, which happens about 80% of the time.

The watchdog monitors which skills are loaded, which decisions are made, & the success rate. Every preflight decision is logged. Every skill invocation is a named, versioned artifact.

Overnight, asynchronous inference³ processes the day’s trail. It decides which new skills should be developed, & which parts of existing skills should become deterministic code. Calendar scheduling is a good example : an LLM should not be comparing free & busy slots ; Rust is much better at that. The system rewrites its skills library & restarts itself in a self-improving loop.

Yesterday was the first day the watchdog did not suggest any improvements. I doubt it will continue. But it hints at something : at some level of improvement, the system reaches a plateau. Only genuinely new exceptions need human help.

That commitment is one quarter of Accenture’s annual labor cost.¹

Three structural models are emerging.

The Balance Sheet. Microsoft & Amazon fund FDE teams from existing headcount. No external capital. Speed & control : Microsoft can reassign engineers without board approval. Salesforce, for example, has committed 1,000 FDE roles.²

The Standalone. OpenAI & Anthropic created standalone entities with external private equity. OpenAI’s Deployment Company raised $4b at a $14b post-money valuation with a 17.5% return floor.³ Anthropic raised $1.5b from Blackstone ($300m), Hellman & Friedman ($300m), Goldman Sachs ($150m), & others.⁴ Scale without diluting the parent. OpenAI acquired Tomoro, a 150-person Edinburgh consultancy with clients including Virgin Atlantic, Tesco, & the NBA.³ Anthropic targets Blackstone’s 275 portfolio companies first.⁴

The Partner Ecosystem. Google Cloud committed $750m to a partner fund rather than building direct.⁵ Capital flows to system integrators & specialists who deploy Google’s models. Leverage : one dollar mobilizes many dollars of partner headcount.

Why now? The bottleneck shifted from model capability to deployment. GPT-4, Claude, & Gemini are powerful enough. Most enterprises cannot install, configure, & operate them without embedded engineering.

FDE investment is a moat. Education builds trust : embedded engineers teach the customer how to use AI. Once a team is trained on one lab’s patterns, retraining on a competitor’s stack is friction no manager volunteers for.

They also see proprietary workflows, data schemas, & failure modes no API call reveals, & that intelligence flows back into model tuning. They expand across the organization, & when a competitor knocks, the embedded team is the defense.

The switching cost is institutional, not technical. And there’s $10b behind it.

Most models sit in the self-expression half of the map, which makes sense given the training data.

Surprisingly, the models are far apart. Gemini 3.1 Flash Lite & Qwen 3.6 Flash sit as neighbors, furthest in self-expression.

GPT-4o & DeepSeek R1 are near-twins, one trained in San Francisco, one in Hangzhou.

DeepSeek R1 & DeepSeek V4 Flash come from the same lab but lie at opposite ends of the secular / traditional axis.

Shared training data & similar labelers explain the near-twins. Different post-training choices explain the strangers. Common Crawl is 46% English², so the base voice a model imitates is a college-educated American online. Anthropic then aligns Claude to principles from the UN Declaration of Human Rights³, a liberal document by construction.

Grok is off on its own, a traditional independent.

This variance changes the shopping list. Every RFP for an enterprise model today scores price, latency, context window, & benchmark scores. Worldview is not on the list. Should it be?

For code generation, SQL, log parsing, & image classification, that is fine. A computer program has no politics.

The moment a model is used for business decisions in a specific market, its worldview is a live input. Marketing copy, predictions of user behavior, & customer support tone all have to match the values of the target demographic.

AI worldviews have never been considered as part of AI procurement, but for certain use cases, it may need to become a consideration.

What matters is the router, a small piece of code that decides which tier of model handles each request. Get the router right & 70-80% of traffic runs on local models that cost nothing per call, or on async models¹ that reduce AI spend by 90%+.

Brian Armstrong made the same point last week about how Coinbase cut AI spend in half while token usage grew², paraphrasing :

How to keep AI spend flat while token usage grows exponentially : not with friction & spend alerts. With better defaults, routing, & caching. Engineers can choose any model they want, but defaults matter.

The routing problem has three layers, and each does a distinct job :

• Skill classifier turns a raw user request into a concrete operation. It answers what the task is. Draft-a-reply, summarize-a-repo, run-a-migration. The classifier is intent recognition.
• Router decides which tier executes the classified operation. It answers which model runs it. The router does not read the prompt. It reads the classifier’s label plus a few features : complexity, context size, historical success rate.
• Model selector picks the cheapest model within a tier that meets a confidence threshold.

Classifier & router are not the same. The classifier is a language problem ; the router is a scheduling problem. Conflating them buries the model choice inside the prompt & kills the ability to A/B different models against the same operation.

Local compute is close to free. Async batch reasoning runs two orders of magnitude cheaper than real-time inference¹. So the real question is narrower : what fraction of work needs real-time answers?

Surprisingly little, once the system can queue work.

Queueing is why this works. A draft reply, a repo summary, a diligence memo, a nightly evaluator run : none of these need to return in a second.

We built the first version of this into our agent runtime. The router already scored tasks on complexity, context size, & local memory retrieval. Two feedback mechanisms now sit on top of the router, & they operate on different time scales :

• Synchronous failure-mode signals. A predictor annotates each incoming route with five features : missing repo context, long dependency chains, risky migrations, security-sensitive prompts, & high-consequence writes.
• Nightly closed-loop feedback. A batch evaluator scores yesterday’s traces overnight & updates the router’s weights, running on async inference on Sail to keep the evaluation cost near zero.

The synchronous predictor catches known-hard tasks before they fail. The nightly loop discovers new failure modes the predictor missed.

Once skill distillation flattens the operation set, 70-80% of agent traffic can run on local models³ for most non-coding work.

The implication : design your system around routing, not around models. Pick your models last.

Two of five public software sectors are up over the last year. The other three are bleeding. The buying pattern is consistent : fund the AI stack, cut everything else.¹

DigitalOcean leads the 1Y board at +430%, with Datadog, Palo Alto Networks, & Fortinet each clearing +50%. At the other end, Monday.com, HubSpot, & Atlassian have struggled.

The sector cuts make the dispersion legible.

Infrastructure & Dev Tools (13 names) leads at +68.5% 1Y, 21.4% revenue growth, 10.0x EV/Sales. DigitalOcean, JFrog, Datadog, MongoDB, Cloudflare, Snowflake, & Confluent sit here. The basket combines agent compute, the AI data stack, & developer tooling.

Security (8 names) is the only other positive sector at +17.6% 1Y, with the highest multiple in the sample at 11.6x EV/Sales & the fastest growth at 24.1%. Palo Alto Networks, Fortinet, CrowdStrike, & Okta carry the basket. AI is expanding both the attack surface & the defense surface, & the market is paying up for the line item that holds the perimeter.

AI & Mega-cap Platforms (9 names) is down -5.9% 1Y despite 21.5% growth. Apple, Microsoft, Nvidia, Meta, Oracle, & Palantir sit here. Growth & profitability are not enough.

Communications & Collaboration (9 names) is down -6.6% 1Y on 8.2% growth & a 2.4x multiple. The one exception proves the rule. Twilio is up +62% because AI agents send messages, & every agent needs a phone number. Even inside a losing sector, the AI-adjacent name wins.

Business Applications (48 names) is the carnage at -36.2% 1Y, 12.5% growth, 3.4x EV/Sales. Salesforce, Workday, & ServiceNow anchor the bucket. This is most of public SaaS, & it is the seat-priced layer most exposed to agent substitution.

Growth does not separate the winners from the losers. Infrastructure, Security, & the mega-caps all grow around 21%. What separates them is category. The market pays premium multiples for sectors CIOs believe are necessary during AI, & punishes horizontal application software where AI threatens the seat-based model.

Marc Benioff told the Logan Bartlett Show in September 2025 :

I’ve reduced it from 9,000 heads to about 5,000, because I need less heads.²

Agentforce now handles half of Salesforce’s customer interactions. Every CIO running Service Cloud just heard that on the same podcast.

The public market is pricing the same question I asked in So You Want to Sell Inference : is your product a payment processor or a software business? Twilio, DigitalOcean, Cloudflare, & MongoDB sit on the token path because AI agents send messages, run on compute, route through edges, & query data.

The rest of the software market trails. The top 1% of companies spend $89k per engineer per year on AI, 40% of a fully-loaded $224k senior engineer salary³.⁴ The median spends $137. That is the gap : 2.3x at the frontier, 0.4x at the top of the market, near zero at the median.

How close does the rest of the market get? Three scenarios bracket the answer.

Bear (token deflation wins), Base (top-1% trajectory tapers), Bull (rest of market reaches Anthropic’s ratio by 2029). Each scenario maps to an annual AI bill per engineer.⁵

In the Bull case, the AI bill alone per engineer matches an entire median-SaaS employee’s revenue contribution.⁶ Anthropic & OpenAI already generate $14m & $6.5m in revenue per employee, the highest in the Forbes Global 2000.⁷

The cost structure follows the revenue structure.

Bull drivers : frontier model prices hold as training costs plateau & demand outruns supply. Agentic workflows consume tokens at orders-of-magnitude higher rates than chat, with Goldman Sachs projecting a 24-fold rise in token consumption by 2030.⁸ If a rival ships features faster, the AI bill stops being optional.

Bear counterweights : token prices have fallen 10x per year for three years.⁹ Open-weight models close the quality gap at a fraction of the cost.¹⁰ Companies that ration usage by role or workload bend the curve.

One of these scenarios will land closer to truth in 2029. Which one are you modeling for 2027?

What if there is no immediate moat? What if the moat is earned?

Leading moats exist at founding. Technical differentiation, a novel architecture, a proprietary dataset. You can point to them in the seed deck. They are most common at the infrastructure layer, where the product is the technology.

Lagging moats are earned through years of execution. Economies of scale, brand, channel relationships, embedded workflows. They cannot be drawn on a competitive landscape matrix because they are not built yet.

Application companies win with lagging moats. Infrastructure companies need leading moats to get off the ground because they require more research & capital to develop.

Salesforce never had a leading moat. Siebel had better technology in 1999. Salesforce won on sales muscle, brand, & a ten-year head start on the cloud-CRM category. Every moat that defends Salesforce today was earned, not engineered.

Snowflake took the other path. The company separated storage from compute when no one else had. That is a leading moat, & it bought the company the runway to build the lagging ones. Marketplace distribution through the hyperscalers. Brand recognition with CIOs. Switching costs embedded in every data pipeline.

Hamilton Helmer’s 7 Powers¹ framework helps separate the two. Scale economies, brand, & switching costs are lagging by construction. They require volume, time, & embedded customers. Counter-positioning, cornered resources, & process power can be leading if you have them at founding, but most app-layer startups do not.

Network economies are earned. Slack, Figma, & GitHub got big before clones could catch them.

Application companies earn their moats over time. Focus, execution, & a market moving faster than incumbents can defend are enough. The moat shows up later. It is no less real.

The honest answer at the app layer : we are building one.