When Andrew Moore, the head of Google Cloud AI, speaks about the future of artificial intelligence, the technology industry listens. In a wide-ranging interview with TechCrunch, Moore laid out what he sees as the three defining frontiers of model capability — areas that will shape how enterprises adopt, deploy, and ultimately profit from AI systems in the years ahead. His framework offers a rare window into how Google, the company arguably most responsible for the modern AI boom, is thinking about the next phase of development beyond the current arms race for bigger and faster models.
Moore’s thesis is straightforward but consequential: the AI industry is moving past the era where raw model size and benchmark performance are the primary differentiators. Instead, he argues, the competitive terrain is shifting toward three distinct frontiers — reasoning depth, multimodal fluency, and agentic capability. Each of these represents a different axis along which models must improve to become genuinely useful for business applications, and each poses fundamentally different engineering and research challenges.
Reasoning Depth: Moving Beyond Pattern Matching to Genuine Problem-Solving
The first frontier Moore identified is what he calls “reasoning depth” — the ability of AI models to engage in multi-step, logically coherent problem-solving rather than simply retrieving and recombining patterns from training data. This is the area where Google has invested most heavily through its Gemini model family, and Moore was candid about the difficulty of the challenge. According to the TechCrunch interview, Moore noted that current models can appear to reason while actually performing sophisticated pattern matching, and that closing the gap between apparent reasoning and genuine logical inference remains one of the hardest problems in AI research.
This distinction matters enormously for enterprise customers. A model that can pattern-match its way through a customer service interaction is one thing; a model that can reason through a complex supply chain optimization problem, weighing trade-offs and accounting for constraints that weren’t present in its training data, is something else entirely. Moore suggested that Google Cloud’s approach involves combining large language models with structured reasoning frameworks — essentially giving models access to formal logic tools that can verify and extend their outputs. The implication is that the next generation of enterprise AI won’t just be about bigger models, but about models that know when to call on external reasoning engines.
Multimodal Fluency: The End of Text-Only Intelligence
The second frontier Moore outlined is multimodal fluency — the ability of models to work across text, images, video, audio, and code with the same facility that current models handle text alone. Google has long positioned itself as a leader in multimodal AI, and the Gemini family of models was designed from the ground up to handle multiple input and output types. But Moore drew a distinction between models that can process multiple modalities and models that are truly fluent across them — able to reason about the relationships between a chart, the text that describes it, and the underlying data it represents, all simultaneously.
This frontier has immediate commercial implications. Industries like healthcare, manufacturing, and financial services generate enormous volumes of data that span multiple formats — medical images alongside clinical notes, sensor data alongside maintenance logs, financial documents alongside market data feeds. Moore told TechCrunch that Google Cloud is seeing the fastest enterprise adoption growth in use cases that require models to synthesize information across these different data types. The companies that can build models with genuine multimodal fluency, rather than bolted-on capabilities, will have a significant advantage in capturing enterprise spending.
Agentic Capability: When Models Start Doing, Not Just Thinking
The third and perhaps most consequential frontier is what Moore called “agentic capability” — the ability of AI models to take autonomous actions in the world, not just generate text or images. This is the area generating the most excitement and the most anxiety across the technology industry. Agentic AI systems can browse the web, write and execute code, interact with APIs, and chain together multiple steps to accomplish complex tasks without human intervention at each stage.
Moore was measured in his assessment of where agentic AI stands today. While acknowledging that Google has made significant progress with its agentic frameworks within Google Cloud, he emphasized that reliability remains the central challenge. An agent that completes a task correctly 90% of the time sounds impressive until you consider that in enterprise environments, a 10% failure rate on autonomous actions could be catastrophic. Moore indicated that Google’s approach focuses on building what he described as “guardrailed autonomy” — systems that can act independently within well-defined boundaries while escalating to human oversight when they encounter situations outside their confidence thresholds.
The Competitive Dynamics Behind Google’s Framework
Moore’s three-frontier framework is not merely an academic exercise. It reflects the intensifying competition among cloud providers to capture what analysts estimate will be hundreds of billions of dollars in enterprise AI spending over the next decade. Microsoft, through its partnership with OpenAI, has been the most aggressive in pushing agentic capabilities through its Copilot products. Amazon Web Services has taken a more infrastructure-focused approach, emphasizing model choice and deployment flexibility through its Bedrock platform. Google, with Moore’s framework, appears to be staking out a position that emphasizes depth and integration across all three frontiers rather than leading on any single one.
The timing of Moore’s public comments is notable. Google Cloud has been on a sustained growth trajectory, with its AI-related revenue becoming an increasingly significant contributor to the division’s overall performance. By articulating a clear framework for where model capabilities need to go next, Moore is effectively setting the terms of the conversation for enterprise buyers who are trying to decide where to place their bets. It’s a classic Google move — using technical authority to shape market expectations.
What Enterprise Buyers Should Take From Moore’s Analysis
For CIOs and CTOs trying to make sense of the AI market, Moore’s framework offers a useful organizing principle. Rather than evaluating AI platforms solely on benchmark performance or model size, enterprise buyers can assess vendors along each of the three frontiers. Does a given platform offer genuine reasoning capabilities, or just sophisticated autocomplete? Can it handle multimodal data natively, or does it require separate pipelines for different data types? And how mature are its agentic capabilities, including the safety and reliability mechanisms that make autonomous action viable in production environments?
Moore also addressed the question of open-source versus proprietary models, a topic that has become increasingly contentious as Meta’s Llama family and other open-weight models have gained traction. He acknowledged that open-source models have made remarkable progress but argued that the three frontiers he identified — particularly agentic capability with enterprise-grade reliability — are areas where proprietary models backed by significant infrastructure investment will maintain an advantage. This is, of course, a self-serving argument from the head of a proprietary cloud AI platform, but it’s one that resonates with enterprise buyers who are wary of the support and reliability guarantees available for open-source alternatives.
The Road Ahead for Google Cloud and the Broader Industry
Moore’s comments come at a moment when the AI industry is grappling with questions about the pace of progress. After several years of dramatic capability improvements driven primarily by scaling up model size and training data, there are signs that this approach is yielding diminishing returns. The three frontiers Moore described represent alternative axes of improvement that don’t depend solely on making models bigger — they require architectural innovation, new training methodologies, and tighter integration between AI models and the systems they interact with.
Whether Google can execute on this vision remains an open question. The company has a long history of producing world-class research that doesn’t always translate into market-leading products. But Moore’s framework suggests that Google Cloud is thinking about AI capability in a more nuanced and ultimately more useful way than the simple bigger-is-better narrative that has dominated the industry conversation. For enterprise buyers, investors, and competitors alike, the three frontiers of reasoning depth, multimodal fluency, and agentic capability offer a roadmap for understanding where the real value in AI will be created — and captured — in the years ahead.
The stakes are enormous. According to recent industry estimates, enterprise spending on AI infrastructure and services is projected to exceed $300 billion annually by 2028. How that spending is allocated will depend in large part on which vendors can demonstrate genuine progress along the frontiers Moore has identified. Google, with its deep research bench, massive computational resources, and growing cloud customer base, is well-positioned to compete. But so are Microsoft, Amazon, and a growing roster of well-funded startups. The race to define the next era of AI capability is far from settled.