You can outsource intelligence, but not understanding

The new rules of growth: 5-7 months to $100M

Muzamil opens with the thesis that anchors Beyond the Bubble: the financial hype around AI is real, but the technology underneath is a genuine step change. Hashim’s first substantive answer sets the frame for everything that follows.

“If you think of the cloud era, companies had like 5 to 7 years to reach 100 million ARR,” Hashim says. “In this AI-native environment, you’re seeing tools like Cursor, Replit, lovable, and others reach that same scale in 5 to 7 months.”

The compression is real, but Hashim refuses to sell it as a template. It creates a second problem: “It’s never been easier to scale, but it’s also more difficult to differentiate yourself.” His prescription is a line he borrows from Lovable: “You don’t need a minimal viable product, you need a minimum lovable product.” Anyone can ship an app in an afternoon. That is precisely why shipping an app in an afternoon is no longer the achievement.

Supernovas, shooting stars, and the wrong goalpost

Muzamil pushes back on the Cursor-and-Lovable narrative directly. His argument is that those companies are outliers riding a wave, and that the kid in the basement who could build a $50,000 business is instead aiming at a $60 billion one and missing both.

Hashim agrees, and reframes the market using Bessemer’s taxonomy. “You have like supernovas which are like your hyper growth,” he says, describing companies that reach 40 million ARR in year one and 200 million by year three. “But you also then have your shooting stars who are more, you know, like typical kind of SaaS companies. They hit 3 million ARR in the first year, then 12, then 40.”

What Google’s team actually works on with founders, Hashim explains, is pulling them out of the hype cycle. He draws on his own experience running Magic Mountain, the fitness app he co-founded and exited: “I was, you know, turning up at gyms, giving out flyers. I was trying to look for key partnerships, as opposed to just spending lots and lots of money just to acquire customers that maybe just shoot off.” The moat, he argues, comes from three places: proprietary data, clear workflow orchestration, and predictable unit economics on tokens.

Tokenomics as the central architectural decision

The tokenomics thread is where the conversation gets sharpest. In the early days of the current cycle, Hashim says, founders picked a model the way people pick a bank, one relationship, one provider. That is no longer viable.

“You maybe don’t need, let’s say, a Fable to edit a deck,” he says, referring to frontier-class models. The work he now does with later-stage startups is essentially forensic: mapping every step of a workflow, identifying which steps need frontier reasoning and which can be handled by a Flash or Flashlight model.

He frames the internal shift plainly: “We need to align the right model to the right workflow, and the right user.” Value max, not token max. This is where Google’s Model Garden inside Agent Platform, formerly Vertex AI, becomes strategically legible, with 200-plus models available inside one governed environment. The orchestration decision, not the model decision, becomes the founder’s real work.

Why Google ships both Gemini and Gemma

Muzamil presses on the strategic logic of Google running both a closed frontier family (Gemini) and an open one (Gemma). What is the economic model? What does Google gain from Gemma the way it gained an ecosystem from Android?

Hashim’s answer is that the two serve genuinely different customers. Gemma sits at “edge and customization,” developers who want to run locally, fine-tune, keep data private, control cost. Gemini sits at enterprise cloud orchestration, multi-step reasoning, multimodal context, safety guardrails, procurement-ready. “The world needs both for different use cases,” he says, “and that’s what Google’s strategy really is, to be the hyperscaler that offers that.”

He points to Google’s Anthropic relationship as evidence. Anthropic models run on Google’s infrastructure, and on some benchmarks faster than they do directly on AWS. The bet is that being the neutral substrate for the entire stack beats being the winner of any single model race. Muzamil connects this back to Sundar Pichai’s 2016 AI-first pivot, which Hashim describes as the origin of a decade-long partner-of-choice posture.

Shadow AI and who owns the training signal

Muzamil raises what may be the most consequential question for enterprise AI adoption: are prompts and workflows being used to train the very models that will eventually disrupt the enterprises typing them in? He notes, without naming names, that at least one major provider does exactly this.

Hashim’s answer is procedural. For a developer prototyping, use Google AI Studio. For anything inside a company account, use Agent Platform. “That is your data, you own it, and you control it,” he says. Shadow AI, employees using unlicensed tools and leaking proprietary data into public models, is the problem enterprise buyers are now solving for first, before they solve for capability.

The point matters beyond Google. Hashim is describing the shape of enterprise AI procurement going forward: sandboxed environments, clear data ownership, and a governance layer that sits above model choice.

Teams of agents, structured like teams of humans

The multi-agent architecture Hashim describes is the operational consequence of tokenomics. Frontier models running 24/7 across sub-agent tasks are, he says, prohibitively expensive: “peak frontier models can be too expensive for those sub-agent tasks.”

The design pattern he keeps returning to is organizational. “The same way that how would you build a human team? You’d have a manager, then you’d have maybe a couple of ICs underneath and everyone specializes in different tasks. That’s exactly the same thing that we should be doing.” A frontier model as the brain, cheaper specialist models as the individual contributors, an orchestration layer routing work between them.

This is where he sees 2025-26 heading: vertical AI, agents deployed for a specific industry, maturing into what he calls AIOS, an agentic operating system connecting disparate tools. And beyond that, physical AI, robotics, factories, agriculture, automotive. “This is real next level super intelligence,” he says.

The data-in problem and why most of the world is stuck

Muzamil offers what may be the single most useful frame in the conversation. AI, he says, is simple: data in, decision, insight out. Everyone is arguing about the middle. Almost no one is solving the first layer. “The agent part is being solved,” he says. “The data in layer just does not exist. It’s because we never really thought about it.”

He describes the geography of the gap directly. San Francisco is “on a different frequency,” operating in 2100 while the East Coast outside New York is still comfortable with chat. In Pakistan and Indonesia, the digital transformation that AI transformation depends on is still in flight.

Hashim accepts the frame. He notes that a surprisingly small share of the world’s data has actually been captured, and that the frontier will move in two directions to address it: localization, through Gemini Live’s language coverage and founders fine-tuning on local linguistics, and synthetic data combined with physical capture, with Waymo cars building new data layers as they drive. But he concedes the gap. “I still think we’re a far removed from that.”

The 50-year question and the case for adaptability

Muzamil closes by asking Hashim to take off his Google hat and answer as the politics, philosophy, and economics graduate he originally trained as. Where does AI take humanity 50 years from now?

Hashim’s answer is deliberately near-term. In five years, humanoids will be more prevalent in daily life, homes will be conversational the way San Francisco homes already are, and physical AI will be embedded in ordinary operation. But his real point is about what humans should hold onto. “We don’t forget the human skills,” he says. “Human empathy, critical thinking, leadership, because that’s what gives us taste, that’s what gives us values, that’s what gives us reasoning and purpose in life.”

He points to Demis Hassabis at Davos telling Dario Amodei that he was “prepared to go slower if it means the world can catch up,” and frames it as a model for how leaders in the space should behave. The optimism he offers is grounded in specific breakthroughs, AlphaFold’s protein structure prediction, the possibility of accelerated clean energy and disease discovery, not the general vibe of progress.

Muzamil closes by naming the three camps he sees: the doomers looking at a 5-15 year horizon of dislocation, the heads-down builders who refuse to think about second-order effects, and the long-horizon maximalists who point to 2,000 years of exponential GDP data and argue that only accelerating through gets humanity to equitable abundance. His own view is that “the timeline is where the problem lies,” and that the work is in lowering the pain of the transition.