Vatsal Bhardwaj

Founder/CEO at Ex-AWS, Meta, MSFT, and Zynga. Talk about generative AI, gaming, building startups.

[Vatsal Bhardwaj @vatsal]

The detail in the Korea Q&A that stood out to me is Next Fest showing demos at random for the first two days before personal recommendations kick in. For a small studio, that means the first 48 hours of demo retention basically decide how much traffic Steam gives you for the rest of the week.

[Vatsal Bhardwaj @vatsal]

@Pirat_Nation If you only tracked the US, UK, Spain, France, Italy and Germany, you'd miss 84.4% of its Steam revenue, per Alinea. China, Taiwan and Hong Kong are 75.9% of the Steam copies.

[Vatsal Bhardwaj @vatsal]

@AravSrinivas Wiring the in-game phone to a real MCP server mid-chase and having Computer rent its own GPU for local Qwen3-TTS is a wild loop. Most game AI demos stop at NPC chat, whereas here the agent is provisioning infra and shipping code while you play.

[Vatsal Bhardwaj @vatsal]

@Dexerto Valve told devs in Korea this week that Steam focuses more on weighting discovery toward the original than on taking down clones. With AI making 1:1 ports this fast to ship, every UGC platform is going to get judged on whether its feed can tell the original from the remake.

[Vatsal Bhardwaj @vatsal]

@DeepThinkLabs Which of those agents needs the most correcting from you? I'd guess the playtesting one, since it has to judge whether something is fun, not just whether it runs.

[Vatsal Bhardwaj @vatsal]

@rewind02 How long did walking every street and fixing things up close take compared with the generation steps? And could Opus make those fixes from your notes, or did you do them by hand?

[Vatsal Bhardwaj @vatsal]

@jackfriks Is that one agent with every tool connected, or a separate one per job? Curious how you keep the customer support side away from the app release access.

[Vatsal Bhardwaj @vatsal]

@svpino Do any of the fixes ship without a person approving them yet, or is it still recommend-only? Curious where you drew that line after a few weeks.

[Vatsal Bhardwaj @vatsal]

@emollick Creative work feels this the most. If you can't see why the agent picked a layout or a level structure, you can't redirect it halfway, you just throw the result out and pay for another run.

[Vatsal Bhardwaj @vatsal]

@Thom_Wolf @eliebakouch Seven incidents turning out to be one misconfigured eval is a strong case for the template. Will the hub tag multi-agent setups separately? The shared work directory one seems likely to come up more as people put several agents on one project.

[Vatsal Bhardwaj @vatsal]

@cursor_ai github replay turning a real PR into eval fixtures is a nice touch. Any plans to do the same for other event sources, so an agent can be tested against what real users actually did?

[Vatsal Bhardwaj @vatsal]

@sama Agents' Last Exam already has a task turning architectural drawings into a 3D model. Has anyone tried Terra inside a game editor like Unity or Godot, where it has to actually play the scene to know if a change worked?

[Vatsal Bhardwaj @vatsal]

@emollick Vancouver at −27.6% in the first week vs −3.4% across the seven US cities is a big gap for the same launch. Did GPT-6 Pro check whether that lines up with where the game actually sold, or is regional sales data too thin to test?

[Vatsal Bhardwaj @vatsal]

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.

[Vatsal Bhardwaj @vatsal]

@akshay_pachaar @bot How often do the four leads hand back work that's wrong? I'd want to know how much checking still lands on you with 14 bots.

[Vatsal Bhardwaj @vatsal]

@IGN Chris Zukowski scraped this month's Next Fest: only 2 of the top 25 demos by peak players disclosed AI, compared with 2,427 of the 6,942 games outside the top 100. Hard to say yet whether the label hurts sales or the AI-heavy games just weren't landing with players.

[Vatsal Bhardwaj @vatsal]

@xikhar Once the Tripo NPCs go in, can the agent also set up navigation and traffic through these streets, or is that where you'd have to step in?

[Vatsal Bhardwaj @vatsal]

Unity says about 98% of its CLI runs between July 20 and October 13 came from coding agents and other automation, so plenty of small teams already have an agent driving the engine. If that agent can also write a simple engine from scratch, what keeps Unreal the default for a two-person team?

[Vatsal Bhardwaj @vatsal]

@AnthropicAI Of the $15.1B jump in first-half revenue, $12.1B came from consumption and $3.0B from subscriptions. Does Claude Design stay usage-priced as it moves into creative work? Creators usually want to know what a finished piece costs before they start.

[Vatsal Bhardwaj @vatsal]

@deedydas If token prices keep getting competed down to cost, app wrappers that mark up usage won't have much margin left. Creative tools probably end up charging for the finished output or publishing rather than the tokens burned iterating.