Vatsal Bhardwaj

Anthropic's S-1 shows first-half 2026 revenue hit $16.3B, up from $1.2B a year ago. Of that $15.1B jump, $12.1B was metered token consumption and $3.0B was subscriptions. Agentic loops are what's driving the meter. In Q3, Claude Enterprise users who didn't use Claude Code spent under $75/month on average, while the top 5% using Claude Code spent ~20x that. Meanwhile, customer-serving compute (cost of revenue) was just $2.4B on that $16.3B in H1 revenue—an 85% gross margin before R&D ($9.7B) and a $2.7B jump in cloud partner fees to AWS, GCP, and Azure. For creative and game tools built on top of these APIs, passing raw token costs straight through to creators gets awkward fast, since they pay for every bad take and broken build too. Charging for the finished output or publishing makes a lot more sense than billing for the tokens burned iterating.

Google and DeepMind ran a proactive agent across 20+ teams for 5 months. Two things bit them: the agent flagged old drafts of documents as contradictions instead of iterations, and teams had no clean way to roll back its edits. In a game or film project, almost every folder is full of cuts, alternates, and old drafts. If an agent doesn't know what's canon and can't be undone in one click, people will turn it off.

Clair Obscur: Expedition 33, per Alinea's estimates: 8.7 million copies sold, plus 4.3 million people who played on Game Pass. Steam was still 58.8% of sales. For a turn-based RPG from a debut studio, letting people try it without buying looks like it grew the audience rather than eating it.

Four friends in Argentina sold over 100,000 pre-orders of Kernel Hearts before launch. In Epic's own thread, the studio director credits putting their character into Fortnite and an Epic Games Store pre-purchase offer. Meanwhile Next Fest is full of devs fighting for wishlists. Making the game is the part AI is speeding up. Getting it in front of players is still the hard part.

Steam Next Fest this week: about 8,000 demos, roughly double June, and over a third disclose AI use. Making a playable demo is no longer the hard part. The scarce things are a real hook, consistent taste, and a way to reach players. AI collapsed the cost of supply. It didn't touch the cost of attention.

Useful numbers. If ~46% of new Steam releases already disclose AI, a "No AI" filter soon hides half the store. Players aren't really asking about provenance, they're asking for curation. Discovery that rewards quality is the harder and better fix.

"AI made a game in 10 minutes" is usually a demo, not a product. A real game still needs state, progression, balance, debugging and iteration. Speed to first playable matters. Speed from first playable to something worth playing matters more.

A model can write a Snake clone in one shot. That doesn't make it a creative partner. Ask for your characters, your art, your mechanics and ten rounds of changes. Then watch what happens: The code breaks. The art drifts. The tenth request forgets the third. Bigger models don't close that gap. Harnesses do.

Most AI game demos optimize for the first 30 seconds. Games have to survive the next 30 minutes. Generate an asset? Easy. Keep mechanics, state and difficulty coherent after 20 iterations? That's the real work. The bottleneck isn't generation. It's orchestration.

We ran a "Bring Your Own Character" game jam last weekend. Over 2,000 participants. The entries that stood out didn't start with a mechanic. They started with a character, a story, an identity - and built the game around it. That order matters. The next wave of games won't begin with an engine choice. It will begin with intent.

Builders are tired of overhyped tools. Not because they’re cynical. But because they’ve seen the pattern too many times. A tool launches with a slick demo. A few early use cases look promising. The homepage is filled with speed claims and screenshots. But two weeks into integrating it into a real workflow, it breaks. Or worse, becomes invisible. The problem isn’t the ambition. It’s the optimization. Too much AI infrastructure is designed for the pitch, not the product. For impressing investors, not helping engineers. For noise, not outcomes. Dashboards don’t help if they only report problems you can’t act on. Token pricing doesn’t matter if context windows force you to trim the value out. Tool integrations are useless if they can’t handle retries, fallback logic, or evaluation at scale. The only three questions that matter for any AI infra tool: → Does it save time? → Does it reveal logic? → Does it remove friction? If the answer is no, it’s not infrastructure. It's interface. This is where a lot of infra teams get it wrong. They assume the job is to make AI “look good.” But the real job is to make it usable at the edges: when prompts fail, when outputs misfire, when context resets mid-flow. Quiet tools win. The ones that disappear into the system and make everything else smoother. That’s the stack people actually keep. So if you’re building infra right now, ask: What real-world pain does this solve at week 4, not day 1? And if you’re using AI tools daily: What’s one that’s actually changed the way you work?

AI can’t replace imagination. But it can speed up play. There’s a recurring question I get: Where is AI most useful in the creative workflow? It’s not ideation. It’s acceleration. At @jabali_ai, we don’t use AI to generate original concepts or replace vision. We use it to compress the time between imagining something and testing it. That matters more than people realize. The gap between “what if?” and “can we try it?” is where most creative momentum dies. Too many good ideas never make it to prototype, not because they were bad, but because they were expensive to explore. Here’s where AI becomes useful: → Drafting narrative branches so creators can explore pacing and structure before writing final dialogue → Assembling modular logic blocks so mechanics can be playtested while design is still fluid → Generating asset variants for early feedback, before production budget gets involved The outcome isn’t fully AI-generated games. It’s a creative workflow where worldbuilding, logic, and visuals evolve together, faster. It’s not about shipping with zero effort. It’s about reducing the cost of iteration until testing ideas becomes part of the creative loop—not a postmortem step. That’s the real leverage: You don’t need better ideas. You need faster ways to prove the right ones are worth building. AI doesn’t replace the creative process. It collapses the timeline between intention and interaction. And that’s where play gets better.