James Hawkins stared at the data pouring into PostHog every day. Session replays. User clicks. Error logs. The kind of granular product insight that most companies would kill for. Yet turning that firehose into proactive intelligence demanded something more than off-the-shelf models. So PostHog decided to train its own.
The announcement landed on May 27. Customers on the US cloud would opt in by default. EU users stay out unless they choose otherwise. Data gets anonymized. Training starts June 29. No selling to third parties. The goal? Smarter session replay analysis at scale. Synthetic user testing before code ships. Predictions that nudge conversion rates higher without burning extra tokens. “I really think we’re on the verge of some of our best work through the next six months,” Hawkins wrote in the PostHog blog post.
Simple on paper. Brutal in practice.
Frontier AI training costs have climbed 3.5 times per year since 2020. They double every seven months. Power demands double annually too. Today’s biggest runs suck down tens to hundreds of megawatts. Enough to light a medium-sized city. Epoch AI documented the surge in its February trends report. Those numbers keep rising through 2030 if current patterns hold.
Yet efficiency gains chip away at the problem. Pre-training compute efficiency improves roughly three times per year. Algorithms get better. Hardware squeezes more from each watt. The net result? Bigger models. Higher absolute spend. And a widening gap between those who can afford the tab and those who cannot.
PostHog sits somewhere in the middle. Not chasing GPT-scale frontiers. Focused instead on narrow, high-value tasks tied to its product analytics core. The company already ships features like an AI installation wizard and PostHog AI assistant. Those rely on existing large language models. The next wave demands models that understand PostHog data deeply. Models trained in-house. On anonymized customer instances. With explicit consent mechanics.
Hawkins laid out the trade-offs in plain language. Opt-out for EU cloud users and anyone bound by BAA or MSA agreements. Everyone else starts opted in. Change your mind anytime in organization settings. Features built on the new models simply won’t appear for those who decline. Enough data matters. Without volume, the models stay mediocre. “Put simply, because otherwise we will not have enough data to train a model that’s actually useful.”
The transparency stands out. Most firms tuck data-use clauses into terms-of-service updates. PostHog emailed customers, pushed in-app notifications, and published the full plan. No data leaves their control. No third-party model providers touch it. The bet rests on turning product usage data into proprietary advantage. Session replays that spot patterns across thousands of users. Synthetic tests that catch friction pre-launch. Behavior models that suggest product tweaks before frustration builds.
This mirrors a broader shift. Companies once happy to call OpenAI APIs now calculate the long-term math. Token costs add up. Context windows limit nuance. Rate limits throttle ambition. Data privacy rules tighten. And the models trained on generic internet scrapes miss the specifics that separate one SaaS product from another.
Numbers tell the story. Training a 70-billion-parameter model ran between $1.2 million and $6 million last year on clusters of hundreds of H200 GPUs. Frontier systems pushed past $100 million when all costs factored in. Some estimates for next-generation runs flirt with billions. A Nvidia executive told Axios in April that compute costs for his team now dwarf employee salaries. “The cost of compute is far beyond the costs of the employees,” Bryan Catanzaro said.
Yet inference prices fall fast. Algorithmic leaps cut costs forty times per year in some benchmarks. Hardware improves. Smaller models deliver surprising punch. DeepSeek claimed a training run under $300,000 for a competitive reasoning model. Others report 45 percent drops in effective training expense thanks to better chips and techniques. The economics split. Pre-training a new foundation model stays punishing. Fine-tuning or training specialized models on proprietary data looks increasingly rational.
PostHog chose the latter path. Its vision centers on PostHog Code, now in beta. A product editor rather than a code editor. One that surfaces answers, acts on them, and improves. The company sees AI moving from helpful assistant to self-driving coworker. That requires understanding not just code syntax but product context. How users actually behave. Where flows break. What changes lift metrics.
Challenges pile up quickly. Data quality varies. Anonymization must preserve signal without leaking identity. Training loops demand constant iteration. Early experiments with third-party models showed promise but hit walls on scale and specificity. So the team builds its own platform. It hires AI researchers. It accepts that some customers will walk away. Better that than quiet changes that erode trust.
The power question looms larger. One gigawatt-scale AI data center can cost $35 billion and take years to build. Largest announced clusters already top 700,000 H100 equivalents. Microsoft’s plans reach millions of GPUs. Electricity demand strains grids. Regulators watch. Investors pour in anyway. More than $170 billion raised by frontier labs in recent years.
And still the efficiency curve bends. Compute efficiency for pre-training doubles roughly every 7.6 months. Chip performance per dollar rises 37 percent annually. Energy efficiency of GPUs climbs 34 percent a year. These gains buy time. They do not erase the trend toward concentration. A few organizations control the largest clusters. Everyone else finds niches.
PostHog’s niche sits inside product development workflows. Teams already trust it with analytics, replays, feature flags, and error tracking. Extending that trust to model training feels like a natural evolution for some. A privacy risk for others. The opt-out default on EU infrastructure reflects regulatory reality. The opt-in default elsewhere reflects data hunger.
Results remain unseen. Training begins at the end of June. Early models will target session replay insights first. Scaling detection of issues that today require human review. Then synthetic testing. Then predictive nudges. Each step tests whether proprietary data yields proprietary performance worth the operational weight.
Other firms watch closely. Some already train on customer data under stricter contracts. Others avoid it entirely, preferring synthetic data or public sources. A few push boundaries and face backlash. The industry sorts itself into camps: those who own their models end to end, those who orchestrate existing ones, and those who simply consume outputs.
Hawkins sounds optimistic. Past AI additions to PostHog simplified the product and increased value. He expects the same here. But he admits the ideas stay experimental. “It will take iteration to figure out how to train models effectively, and what data is actually useful.”
That humility matters. AI development at any scale rewards patience and clear-eyed accounting. Costs rise. Capabilities rise faster in some dimensions. Power bills arrive monthly. Talent demands equity or seven-figure packages. The organizations that survive will treat training decisions as strategic bets, not technology experiments.
PostHog placed one such bet this week. Transparent. Customer-centric. Focused on concrete product outcomes rather than vague intelligence explosion. Whether the models deliver remains the test. The data exists. The compute exists. The question is whether the synthesis creates something customers cannot live without.
Plenty of companies face the same calculation in 2026. Build your own. Pay per token forever. Or fall behind. The middle ground shrinks. PostHog chose a side. Others will follow. Some will regret it. A few will point to this moment as the decision that changed their trajectory.
WebProNews is an iEntry Publication