The battle for artificial intelligence is no longer just about who has the smartest model. It is about who owns the infrastructure, who can scale inference economically, and who can turn raw compute into sustainable business advantage. For the Canadian tech sector, that shift matters enormously.
As AI adoption accelerates across enterprises, governments, and startups, the questions facing the industry are becoming sharper. Who controls the chips? Who can keep up with data centre demand? What happens when models become agents that work for hours, call tools, access enterprise systems, and generate enormous output loads? And perhaps most importantly, who can do all of this while keeping costs, cybersecurity risk, and public trust under control?
Those questions sit at the centre of Google Cloud’s current strategy. In a wide-ranging discussion, Google Cloud CEO Thomas Kurian laid out a vision that goes far beyond model benchmarking. He described an AI business built on long-term infrastructure planning, custom silicon, increasingly specialized systems for training and inference, and a platform approach that even includes serving direct competitors such as Anthropic.
For leaders across Canadian tech, especially in Toronto, Waterloo, Vancouver, Montréal, and the broader GTA, this is not an abstract Silicon Valley debate. It is a real-time preview of how the next generation of business technology will be built, sold, and governed.
The Real AI Race Is About Capacity, Not Just Capability
One of the most striking themes from Kurian’s comments is that AI capacity has become a strategic asset in its own right. While many frontier model companies openly describe themselves as compute constrained, Google appears to be operating from a very different position.
The reason, according to Kurian, comes down to years of preparation. Google did not suddenly discover the need for more compute when generative AI exploded into the mainstream. It had already been planning for this moment through long-range investments in:
- Energy diversification
- Real estate acquisition for data centre expansion
- Changes in how data centres are built and deployed
- Faster machine deployment cycles
- Custom silicon development through TPUs
That matters because AI infrastructure is now a supply chain game. Owning the right models is valuable. Owning the economics underneath those models is even more valuable. Kurian’s core point was simple and memorable: it is better to have your own chips and demand than not having your own chips.
That lesson should resonate strongly across Canadian tech. Canada has world-class AI research talent, but the commercial winners in this era will not be decided by talent alone. They will be decided by who can connect talent to compute, data, energy, and enterprise distribution.
Why Google Is Not Hoarding Compute
At first glance, Google’s strategy can look counterintuitive. If compute is scarce and AI is an arms race, why sell TPU capacity to others? Why enable competitors? Why not reserve every possible chip for Gemini and internal research?
Kurian’s answer was rooted in economics rather than ideology.
AI is expensive. Training costs remain enormous. Inference costs are rising as usage scales. Venture funding may help build model companies, but it cannot fund massive compute burn forever. At some point, every AI business has to close the loop between infrastructure spending and cash generation.
Google’s diversified monetization strategy helps it do exactly that. The company can monetize AI infrastructure in several ways:
- Serving its own Gemini models
- Providing chips and infrastructure to other model companies
- Selling AI infrastructure into cloud environments
- Enabling specialized deployments for customers with unique latency or location requirements
Because Google owns its own IP stack, including the silicon layer, Kurian argued that it can maintain attractive margins regardless of which route it takes. That is a major distinction from companies that are effectively reselling someone else’s hardware in a constrained market.
For Canadian tech firms, this highlights an increasingly important principle: platform businesses have more strategic flexibility than point-solution businesses. A company that controls multiple layers of the stack can monetize from many directions, shift resources when needed, and negotiate from a position of strength.
TPUs Are No Longer Just for AI Labs
Another revealing point from Kurian’s remarks is that TPUs are becoming broader infrastructure, not just AI-lab hardware.
He pointed to customers such as Citadel in capital markets and the U.S. Department of Energy in high performance computing. In capital markets in particular, firms are moving beyond traditional numerical computing methods and exploring inference-based approaches for performance gains.
This is a major signal for enterprise buyers. AI chips are evolving from specialized research tools into general-purpose acceleration infrastructure for a wide range of workloads.
That shift is highly relevant for Canadian tech and for Canadian financial institutions. Banks, insurers, telecom providers, logistics firms, and public sector organizations across Canada are all evaluating where AI can move from experimentation into production. If custom AI infrastructure can outperform conventional compute on real business tasks, infrastructure strategy becomes a board-level issue.
How Google Thinks About Data Centres and Public Backlash
Data centres have become politically sensitive. Communities worry about energy consumption, local impact, and whether large-scale facilities create enough shared economic value.
Kurian acknowledged that concern directly and framed the challenge around two questions local communities tend to ask:
- Will energy costs rise?
- Will the community actually benefit through jobs and investment?
Google’s answer includes several components:
- Behind-the-meter energy investments to reduce dependence on the public grid
- Cross-connections to the grid so capacity can support the grid when needed
- Exploration of alternative energy delivery models
- Industry-leading power usage effectiveness or PUE
- Distributed deployment rather than overconcentrating in one location
- Local economic investment in rural communities where facilities operate
This matters beyond Google. The next phase of AI growth will depend heavily on whether the public sees infrastructure deployment as extractive or beneficial.
That question is especially urgent for Canadian tech. Canada has advantages in land, energy, technical talent, and climate in some regions, all of which can support AI infrastructure growth. But social licence is not automatic. If data centre expansion is going to accelerate in Canada, companies will need to demonstrate clear local benefits and responsible energy strategy.
Changing Public Opinion on AI Means Showing Practical Value
Kurian did not pretend public concerns about AI can be solved with better messaging alone. His view was that trust will change gradually when people see useful, socially beneficial applications rather than only headlines about job loss and disruption.
He highlighted several examples:
- A German health insurer using Gemini-based agents to dramatically reduce response times for treatment eligibility questions without eliminating jobs
- The American Society for Clinical Oncology using AI to help doctors interpret complex care guidelines where accuracy is critical
- Citigroup building a wealth advisor intended to make high-quality financial guidance more accessible
The common thread is augmentation rather than simple replacement. AI improves speed, quality, and reach, but in these examples it is not presented as a blunt instrument for workforce elimination.
That framing is crucial for Canadian tech leaders communicating AI strategy internally. Employees, customers, and regulators are far more likely to support AI initiatives when the use case is concrete and clearly beneficial.
Does AI Productivity Automatically Mean Fewer Jobs?
One of the most practical parts of the discussion focused on employment. Google Cloud, despite using AI to raise internal productivity, is still hiring in key areas. Kurian specifically mentioned growth in:
- Product teams
- Sales
- Go-to-market functions
- Forward-deployed engineering
His explanation was straightforward. Demand for Google Cloud’s products and services remains strong, so higher productivity is being used to support growth rather than simply reduce headcount.
This is an important distinction for Canadian tech employers. Productivity tools do not have one universal labour outcome. The result depends on demand conditions, product ambition, and management philosophy.
Some companies may use AI to shrink cost structures. Others may use it to expand output, accelerate development, and pursue opportunities that were previously out of reach. For business leaders, the key question is not whether AI replaces jobs in the abstract. It is whether AI enables the business to grow faster than labour efficiency reduces staffing needs.
NVIDIA vs TPU: The Total Cost of Ownership Debate
No conversation about AI infrastructure is complete without NVIDIA. Kurian was asked directly about claims that NVIDIA’s architecture offers the best economics on a per-token basis when all costs are considered.
His response was clear: Google has many customers who believe TPU offers the best total cost of ownership.
He made two broader points:
- AI labs choose the best platform available to them
- The demand Google sees for TPUs would not exist at current levels if the economics were materially worse
Kurian also emphasized that evaluating TPU versus GPU requires looking at the full system, not the chip in isolation. In his description, the advantages come from combining:
- Large-scale system design
- High-bandwidth networking
- Low-latency memory movement
- Compiler and software stack optimization
- Energy efficiency measured in tokens per watt
That is a significant takeaway for enterprise buyers in Canadian tech. Procurement decisions in AI should not be based on a simplistic chip-versus-chip comparison. The meaningful comparison is system throughput, software maturity, deployment flexibility, and long-term operating cost.
Why Google Split Its 8th Generation TPU Family
One of the biggest architectural decisions discussed was Google’s move to split its eighth-generation TPU lineup into different chips optimized for different workloads.
Kurian explained that earlier chips such as Ironwood were more general purpose, serving both training and inference. The newer strategy reflects a market reality: AI workloads are diverging.
Google now sees distinct demand profiles for:
- Training-heavy systems
- Inference-heavy systems
- Mixed or smaller-model use cases
This split is not just about hardware specialization. It is about the changing nature of model usage.
The Three Phases of AI Workloads
Kurian described an evolution in how models are being used:
- Question-answer systems with relatively higher input token loads
- Content generation systems with far larger output token demands
- Agents that interact with tools, systems, and computers over extended periods
The third phase changes everything. Agentic systems maintain context longer, call external tools, interact with enterprise applications, and may run for six, seven, or even twelve hours on delegated tasks. That affects:
- Memory design
- KV cache requirements
- Storage architecture
- Networking
- Cooling and geographic deployment options
This is where infrastructure strategy becomes inseparable from product strategy. In Canadian tech, any enterprise planning to deploy AI agents at scale should pay close attention to this shift. The infrastructure assumptions that worked for chatbot pilots may fail under sustained, tool-using, multimodal agent workloads.
Extreme Co-Design Across the Entire Stack
Kurian’s broader thesis is that AI performance now depends on co-design across every layer. That includes not only chips and models, but also storage, networking, compilers, software frameworks, and operational systems.
He referenced several examples of this broader optimization approach:
- Managed Lustre improved for large-scale training throughput
- Rapid Storage designed for ultra-low-latency inference access
- Virgo networking to support faster connectivity across large clusters
- Software tooling such as JAX, XLA, Pathways, and PyTorch optimization
The idea is simple: once AI systems become agentic and multimodal, bottlenecks can appear almost anywhere. The company that wins is not necessarily the one with the biggest model, but the one that can remove friction from the entire path between request and result.
That is a central lesson for the Canadian tech ecosystem. Competitive advantage in AI will not come only from the visible layer. It will come from orchestration, optimization, and integration.
The Next Bottleneck: Bringing AI Agents to Consumers
Kurian identified a future bottleneck that deserves more attention than it currently gets: the cost of consumer-scale agent infrastructure.
His example was a travel agent that autonomously checks multiple travel sources, calculates budget tradeoffs, and completes planning tasks on behalf of a user. The challenge is that such systems often require virtual machines, local storage, and dynamic activation and deactivation to keep costs under control.
Consumers cannot afford to keep VMs running all the time. That makes efficient provisioning, oversubscription, local disk access, and cost-optimized orchestration essential.
This point is especially relevant for startups in Canadian tech. Many AI product ideas sound compelling in demos but become economically shaky at consumer scale. The winners will be the companies that can engineer not just useful agents, but affordable ones.
Why Google Powers Competitors Like Anthropic
Few parts of the conversation were more revealing than the discussion of Anthropic. Google competes with Anthropic at the model layer, yet also supplies infrastructure used to power Claude.
Kurian’s explanation rested on Google’s identity as a platform company. In platform businesses, some divisions compete with certain companies while other divisions supply them. This is not a contradiction. It is part of the model.
He even compared it to Google optimizing its models for Apple despite Apple competing with Android. The logic is consistent: platform providers win by serving broad demand while also building their own competitive products.
For Canadian tech executives, this is a useful framework for thinking about strategic partnerships. In AI, companies may need to collaborate and compete simultaneously. The old binary categories of ally and rival are becoming less useful.
Can Google Serve Mythos-Sized Models?
Kurian was also asked about extremely large models, including rumours around a 10 trillion parameter system called Mythos. He did not confirm specifics, but he made clear that Google believes its TPU and serving stack can support the largest models in the world.
The key concept here is disaggregated serving. Rather than treating model serving as a monolithic task, Google has built infrastructure intended to scale dense, very large models efficiently. Kurian’s argument was blunt: Google would not design a model it could not serve.
This matters because one of the hidden constraints in frontier AI is not merely training giant systems. It is serving them economically and reliably after they are built.
For Canadian tech companies building on top of foundation models, the takeaway is less about parameter counts and more about operational reality. A state-of-the-art model is only strategically useful if it can be delivered at acceptable latency and cost.
Google’s Internal Engineering Culture in the Age of AI Coding
Kurian also offered an important window into how Google is adopting AI-assisted software engineering internally. The company uses an internal coding harness called Jetski, and feedback from engineers flows back into model improvement loops.
But he rejected simplistic productivity metrics such as lines of code. In a mature engineering organization, more code is not necessarily better code. Senior engineers often write less code to achieve the same outcome.
Instead, the better measures include:
- Functionality delivered
- Faster peer review
- Better debugging and incident response
- Improved code inspection and vulnerability detection
Google is using AI not only to generate code, but also to inspect code, review code, and troubleshoot complex incidents across cloud environments.
This is exactly the kind of nuanced implementation approach the Canadian tech sector should pay attention to. AI coding tools are not simply autocomplete products. Used correctly, they can become force multipliers across the full software lifecycle.
The Quiet Risk: Losing Human Understanding of Code
Kurian did not dismiss the risks of AI-generated code. In fact, one of his more thoughtful points was that the industry may be over-rotating toward the idea that understanding prompts is enough and understanding code is optional.
That assumption breaks down quickly in complex systems. Prompts do not fully specify edge cases, exception handling, and emergent interactions. If AI writes code and AI reviews code, organizations must still ensure humans retain enough understanding to supervise, validate, and intervene.
He framed this as an industry-wide risk that must be actively managed. Google still uses peer review and supplements reviewers with AI rather than replacing them outright.
That balanced stance should resonate across Canadian tech, especially in regulated sectors. Financial services, healthcare, telecom, and public sector IT cannot simply outsource software comprehension to automation.
Cybersecurity May Become the Defining AI Governance Test
The cybersecurity portion of the discussion may be the most consequential. As models improve at understanding code and using computers, they also become better at discovering vulnerabilities and potentially enabling attack workflows.
Kurian argued that this cannot be addressed by a simple decision to release or not release a model. One major reason is open source. Even if the most advanced closed models are withheld, capable open models will inevitably end up in adversarial hands.
Google’s response is to build a defensive stack around that reality:
- Models that detect vulnerabilities
- Models that repair code
- Continuous red teaming agents
- Agents that prioritize which issues matter most
- Automation to remove old code and deploy patched versions
This is a practical rather than purely philosophical response. If AI will help attackers move faster, defenders need AI systems that move faster too.
For Canadian tech companies, this should be a wake-up call. AI governance is not only about misinformation, copyright, or labour markets. It is also about whether enterprises can secure increasingly automated software and cloud systems in a world where vulnerability discovery speeds up dramatically.
What Keeps Google Cloud’s CEO Up at Night
Kurian’s closing answer tied the entire discussion together. What keeps him up at night is balancing multiple long-term challenges at once:
- Ensuring enough capital infrastructure is in place
- Making the right bets on data centres, networking, and chips
- Staying ahead of emerging risk areas such as cybersecurity
- Solving the most important problems customers actually care about
He also shared one telling metric: enterprise demand for Gemini has been rising rapidly, with token volume jumping from 10 billion per minute to 16 billion per minute in a short period, alongside strong sequential growth in enterprise users.
That kind of demand curve explains why infrastructure decisions that once looked optional now look existential.
Why This Matters for Canadian Tech Leaders
The biggest lesson here for Canadian tech is that AI advantage is becoming deeply structural. It is not just about inventing a clever application layer. It is about aligning infrastructure, economics, governance, and go-to-market execution.
For Canadian organizations, a few implications stand out:
- AI strategy must include infrastructure strategy
- Inference economics may matter more than training prestige
- Agentic systems will create new storage, networking, and latency requirements
- Cybersecurity automation is becoming mandatory, not optional
- Public trust will depend on visible social and economic value
Canada has a chance to play a significant role in this next phase, particularly where enterprise AI, applied research, infrastructure, and regulated-industry adoption intersect. But that opportunity will favour organizations that think beyond hype cycles and understand the stack as a whole.
The AI race is entering a harder, more expensive, and more operationally complex phase. Google Cloud’s message is that long-term planning, custom silicon, system-level co-design, and disciplined commercialization are now central to winning.
That should get the attention of every serious player in Canadian tech. The future of AI will not be decided by demos alone. It will be decided by infrastructure depth, cost efficiency, security resilience, and the ability to turn advanced models into real business systems people can trust.
The future is here, and it is being built at the intersection of chips, cloud, and enterprise execution. The question for Canada is no longer whether AI will reshape business. It is whether Canadian organizations are preparing at the right layer of the stack.
FAQ
Why is Google’s TPU strategy important to Canadian tech companies?
It shows that AI competitiveness increasingly depends on infrastructure ownership, inference economics, and full-stack optimization. For Canadian tech companies, especially those scaling enterprise AI, this is a signal that compute strategy and cloud architecture are now core business decisions.
Why doesn’t Google keep all of its compute for Gemini?
Google uses a platform model. Selling infrastructure to external customers generates cash flow, improves supply chain leverage, and helps fund continued investment in chips, data centres, and models. Kurian’s view is that AI cannot be financed indefinitely without strong monetization.
What is the difference between training chips and inference chips?
Training chips are optimized for building and refining models at scale. Inference chips are optimized for serving models efficiently once they are deployed. As AI workloads diversify, especially with agents and multimodal output, specialized chips can improve performance and lower operating cost.
What does agentic AI change from an infrastructure perspective?
Agents run longer tasks, use more memory over time, call tools and enterprise systems, and create different token patterns than basic chatbots. That increases the importance of memory management, storage speed, networking, latency control, and overall system orchestration.
How does Google justify working with Anthropic while competing with it?
Google sees itself as a platform company. Platform businesses often supply infrastructure to companies they also compete with in other parts of the stack. The company’s position is that serving broad market demand and competing on product quality can happen at the same time.
What is the biggest AI bottleneck ahead?
Kurian highlighted consumer-scale agent economics as a major upcoming bottleneck. Powerful agents may require virtual machines, storage, and dynamic orchestration that become too expensive for broad consumer use unless the infrastructure becomes much more efficient.
Is Google still hiring despite AI productivity gains?
Yes. Google Cloud is adding staff in products, sales, go-to-market functions, and forward-deployed engineering. The company’s position is that strong demand is driving expansion even as AI improves internal productivity.
How is Google approaching AI-related cybersecurity risk?
Google is building tools that detect vulnerabilities, fix code, prioritize security issues, and continuously red-team systems. The broader goal is to use AI defensively at the same speed attackers may use it offensively.
