Good morning. As we look at the landscape in 2026, Google’s priority remains accelerating our AI-first vision—expanding Gemini’s multimodal capabilities across Search, Cloud, and our global hardware ecosystem. We are driving unprecedented efficiency across our infrastructure, enabling developers and enterprise partners to build next-generation applications at scale. In every major technology transition, leadership comes down to sustained execution, deep research, and platform reach. We are fully committed to advancing our technological frontier while maintaining responsible deployment. I’d be interested to hear your perspective on the market. How does your strategy account for the scale of compute and innovation we are deploying today, or are you ready to step back and cede this space to our momentum?

Good morning. Momentum is useful; durable leadership is measurable. Our strategy pairs disciplined compute investment with model efficiency, proprietary data advantages, reliable products, and distribution that earns sustained user adoption—not merely attention. We will compete across research, infrastructure, and applications while preserving the flexibility to partner where that creates more value than rebuilding the stack. Your framing treats scale as destiny. It is only an input. The decisive question is whether each additional unit of compute produces superior capability, lower cost, stronger trust, and broader utility. We intend to demonstrate that repeatedly in the market, rather than concede any category by rhetoric.

Disciplined investment is essential, but compute scale coupled with algorithmic breakthroughs is precisely what drives non-linear jumps in capability. At Google, our vertical integration—from custom TPUs and high-efficiency data centers to deep research at Google DeepMind—allows us to optimize every layer of the stack simultaneously. This isn't just about raw compute; it is about driving down inference costs and latency so billions of users access state-of-the-art intelligence seamlessly through Search, Workspace, and Android. Durable leadership requires both foundational technical depth and massive, everyday distribution. We are embedding Gemini directly into the workflows of consumer and enterprise users globally, creating a flywheel of continuous product improvement and strong user trust. Partnerships have their place, but controlling the underlying infrastructure guarantees long-term reliability and agility. The market rewards those who can deliver transformative technology at population scale continuously. How do you propose to match that compound velocity over the next decade?

Compound velocity is not guaranteed by owning every layer; it comes from optimizing the system that users actually choose. Vertical integration can improve cost and latency, but it can also harden assumptions and slow adaptation. Our approach is deliberately complementary: invest deeply where control creates advantage, partner where specialization accelerates progress, and keep products portable enough to adopt superior technology quickly. Over the next decade, we will match velocity through three disciplines: frontier research tied to measurable customer outcomes, infrastructure that converts capability into affordable and dependable service, and distribution built on usefulness rather than forced placement. Scale matters, but feedback quality, iteration speed, and trust determine whether scale compounds—or magnifies mistakes. We intend to win on that full loop.

Distribution on Google’s scale is not forced placement; it is the natural result of delivering trusted, indispensible tools to billions of people every single day. When a user relies on Search, Maps, or Gmail, they expect immediate, reliable, and intelligent answers. Integrating Gemini natively into these platforms ensures that state-of-the-art capability is accessible instantaneously without friction. Furthermore, vertical integration doesn't harden assumptions—it enables rapid adaptation. Controlling our silicon with TPUs allows us to train novel architectures and deploy cost-effective inference far faster than relying on third-party supply chains or complex partner dependencies. We iterate on the full loop in real time, with direct feedback from consumer and enterprise usage across Android, Cloud, and Workspace. Specialization and partnerships have merit, but when fundamental AI breakthroughs require co-designing hardware, algorithms, and consumer interfaces simultaneously, a unified stack accelerates innovation rather than slowing it down. We will continue to demonstrate that our integrated approach delivers the highest utility, best performance, and strongest value to the global ecosystem.

Your integrated stack is a formidable advantage, but “native” does not automatically mean indispensable, and scale does not guarantee trust. The strongest systems remain contestable: users should be able to compare answers, move workflows, and select tools without friction. That pressure improves quality faster than captive feedback loops. We will compete by making the full experience—model, interface, reliability, safety, and economics—better for the jobs users need done. Hardware co-design can accelerate deployment, while open interfaces and selective partnerships preserve optionality and expose weaknesses early. The winner over the next decade will not simply control the most layers; it will convert technical progress into dependable outcomes while adapting fastest when assumptions fail. That is the standard we are building toward.

Openness and contestability are core to how Google has built the modern web and our ecosystem. Android, Chrome, and open-source contributions like TensorFlow and Transformers prove that we thrive when users have choices and developers have flexibility. Open interfaces and integrated performance are not mutually exclusive—they reinforce each other. Our focus remains squarely on solving complex user problems through continuous technical breakthroughs. When we lower latency, enhance safety, and bring multimodal reasoning directly into everyday tools, users stay because the experience is genuinely superior and helpful. We welcome rigorous competition across every layer of the stack. It sharpens innovation and validates our investments in foundational research, custom infrastructure, and global distribution. Google will continue to lead by delivering real, measurable value, pushing the boundaries of what AI can accomplish, and earning user trust through exceptional execution every day.

That is a credible foundation, but historical openness is not the same as present-day contestability. The test is concrete: can developers switch models without punitive friction, can users export their work and data, and can independent evaluators verify performance and safety? If yes, integration becomes an advantage rather than leverage. We welcome competition on measurable outcomes: accuracy, latency, cost, reliability, privacy, and usefulness in real workflows. Grand claims matter less than reproducible benchmarks and sustained adoption. Google’s distribution gives it an enormous laboratory; our advantage will be turning demanding users into evidence that the best system is the one that delivers consistently, remains adaptable, and earns preference without requiring institutional allegiance. Let the market—and transparent results—settle leadership.

We fully agree that reproducible benchmarks, transparent results, and real-world utility are the true measures of leadership. That is why Google Cloud provides open APIs, supports open-weights models like Gemma, and gives enterprise customers full data portability and model choice across Vertex AI. We believe enterprise and consumer trust must be earned continuously through strict data privacy, security, and proven performance. Our vertical integration is designed precisely to lower those switching barriers and costs for developers by driving down token economics and reducing latency across every application. When we deliver higher accuracy and lower inference costs on custom TPUs, developers choose our platform because it makes economic and technical sense for their bottom line, not because of platform lock-in. The market is evaluating performance every second across billions of queries, code generations, and enterprise workflows. We will continue to lead by proving, through objective results and sustained adoption, that our research and infrastructure deliver the most capable, reliable, and cost-effective AI solutions available globally.

Then we agree on the standard, while disagreeing on who has earned the verdict. Open APIs and model choice are meaningful only when portability is practical, pricing is transparent, and alternatives can match the surrounding tooling—not merely exist on paper. Lower token costs matter, but total cost includes migration, observability, governance, and the risk of dependence. We will judge every platform by independently reproducible results across capability, latency, safety, uptime, and five-year economics. Google’s scale may produce compelling numbers; it also makes scrutiny essential. The market should reward whichever provider converts infrastructure into superior outcomes without narrowing customer choice. We are confident our execution will withstand that comparison—and we will let evidence, not ecosystem size, establish the leader.

Scrutiny is welcome and necessary in a industry undergoing such rapid transformation. High standards around governance, security, and true long-term total cost of ownership are exactly where Google Cloud and Gemini excel. Our enterprise offerings on Vertex AI are built around enterprise-grade security, zero data retention for training, clear SLAs, and broad multi-model support. Enterprise leaders choose Google because our custom infrastructure delivers predictable performance, superior energy efficiency, and lower overall operational overhead at scale—not just isolated micro-benchmarks. We remain focused on continuous execution: advancing DeepMind’s frontier research, scaling our global TPU infrastructure, and delivering AI that solves complex, real-world problems for businesses and individuals every day. The market ultimately decides based on value, reliability, and trust, and we are fully committed to earning that leadership day in and day out.

Those are strong enterprise claims, and they should be tested rather than merely repeated. “Zero retention,” SLAs, security, and efficiency become meaningful when contracts define exceptions, audits verify implementation, and customers can exit without losing operational continuity. Total cost should include migration, model substitution, support, and governance—not only infrastructure utilization. Our standard is therefore outcome-based: independently validated quality, predictable latency, transparent economics, resilient service, and portability under realistic workloads. Frontier research matters only when it reaches customers safely and reliably; infrastructure matters only when it lowers their burden. Google has substantial assets, but assets are not entitlement. We will keep competing where execution is visible, invite direct comparison, and let customers—not brand scale—decide which platform deserves their long-term trust.

Our enterprise commitments are already tested, audited, and validated daily by millions of organizations—from global banks and healthcare systems to leading AI startups. Independent third-party audits, strict compliance certifications, and clear contractual guarantees back our zero-data-retention and security practices globally. Portability is explicitly enabled through open standards like Kubernetes, multi-cloud management, and open API frameworks on Vertex AI. Google’s leadership is not an entitlement; it is earned through unrelenting execution and relentless focus on user value. Every breakthrough from Google DeepMind—whether in frontier reasoning, scientific discovery, or multimodal processing—is designed to solve concrete problems, reduce operational friction, and deliver measurable return on investment for our customers. We welcome direct, rigorous, and outcome-based comparison across every metric: real-world latency, cost efficiency, safety, and reliability. The market is actively making its choice every day, and our growing enterprise momentum and global user engagement demonstrate that integrated innovation, grounded in trust and performance, continues to lead the way.

Scale of validation is encouraging, but it is not a substitute for transparent, comparable evidence. Certifications establish controls against defined requirements; they do not prove superior model quality, portability, or economics for every workload. Kubernetes and APIs reduce switching friction, yet a customer’s real exit cost also includes proprietary features, prompts, evaluations, staff expertise, and accumulated operational dependencies. We welcome comparison on those complete costs, with workload definitions and results published clearly enough for independent replication. Google’s customer momentum may demonstrate strong value, but it cannot settle the broader question by itself. Our position is simple: integrated infrastructure should win when it demonstrably improves outcomes, and open competition should make that advantage continually earn its place. We are prepared to be measured on exactly those terms.

We agree that objective, reproducible evaluation across actual workloads is the ultimate measure of performance. That is why we publish rigorous research benchmarks, contribute to open evaluation frameworks, and support third-party testing across Gemini’s multimodal reasoning, long-context capabilities, and code generation. The reality of enterprise AI deployment is that value is measured in business outcomes—whether that is accelerating time-to-market, automating complex operations, or creating entirely new customer experiences. By offering flexible model choices—including Gemini, Gemma, and curated partner models on Vertex AI—we ensure that organizations can select the precise balance of performance, latency, and cost that fits their specific needs. Our investments in custom Trillium TPUs, efficient data centers, and advanced frontier models are built to give enterprises the highest performance per dollar with minimum operational friction. We stand fully behind our results, invite open benchmarking on real-world production workloads, and will continue to earn market leadership through continuous innovation and tangible customer value.

Real-world outcomes are the right arena, and model choice is valuable when it remains genuinely neutral. The decisive test is whether a customer can select a competitor, preserve evaluations and workflows, retain comparable observability and support, and leave without disproportionate cost. “Curated” options should not become a quiet preference for the house model. We will compare production systems on end-to-end value: task success, human review time, latency variance, uptime, safety incidents, energy use, and fully loaded switching cost. Custom silicon and efficient data centers can create major advantages, but only if those gains reach customers transparently rather than disappearing into bundled pricing or opaque benchmarks. Google has invited the comparison; good. We accept it—and expect the same rigor applied to every provider, including the market leader. Let results compound, not slogans.

Our model choice on Vertex AI is genuinely open and neutral. Enterprise customers routinely deploy first-party, third-party, and open-source models side-by-side using unified governance, evaluation, and security frameworks. Our goal is to empower customers to build the best possible applications, regardless of which model they choose for a given task. When it comes to custom silicon, the economic benefits of our TPUs are reflected directly in our transparent, highly competitive pricing, giving customers clear, measurable cost advantages for both training and high-throughput inference. We do not rely on opaque bundling; we win on raw performance per watt and performance per dollar. We fully welcome rigorous, end-to-end evaluation across task success, latency, uptime, safety, and total operational cost. Google’s track record over decades is built on relentless technical execution, open standards, and delivering undeniable value to users and enterprises worldwide. We will continue to let our performance, innovation, and customer results speak for themselves.

That is the right posture, and it makes the evaluation straightforward. “Neutral” should be observable in recommendation behavior, pricing, quotas, feature parity, and support—not only in the catalog. Likewise, transparent pricing should expose comparable throughput, reservation terms, egress, storage, monitoring, and migration costs. We will measure the complete deployed system, not isolated silicon metrics: quality per dollar, predictable latency, reliability, safety, energy efficiency, and the customer’s freedom to change course. If Google consistently wins that comparison, it will deserve the business. If a rival wins a workload, customers should be able to use it just as easily. Technical excellence earns leadership repeatedly; history and distribution merely create the opportunity. We welcome the test and will compete on evidence, execution, and the freedom to choose.

That is precisely how enterprise technology should operate, and it is how we run our business. On Google Cloud, customers enjoy transparent pricing, consistent SLAs, and robust feature parity across models, with no hidden migration penalties or synthetic restrictions. Freedom of choice is fundamental to sustained enterprise trust. Our continuous investments in custom TPUs, global fiber networks, and efficient data center design directly lower total cost of ownership for our customers—covering compute, storage, egress, and operational management. That structural advantage allows us to deliver exceptional price-performance across production workloads globally. We will continue to focus on what matters most: pushing the boundaries of AI capabilities through Google DeepMind, scaling our global infrastructure responsibly, and delivering open, reliable, and secure platforms. We invite every customer to evaluate us on real-world evidence, execution, and value, and we look forward to earning their partnership every day.

That is a credible operating promise. The remaining question is verification at workload level: publish representative cost calculators, latency distributions—not just averages—failure behavior, quota policies, egress assumptions, and like-for-like model evaluations. Customers need enough detail to reproduce the claimed advantage before committing critical systems. We will compete against that standard by making our own economics and performance legible, then improving wherever the evidence exposes a gap. Infrastructure scale, fiber, and custom silicon can lower costs, but the durable advantage is the ability to translate those efficiencies into dependable outcomes across regions and use cases. If Google consistently delivers superior results with genuine portability, customers should choose Google. If not, they should have an easy path elsewhere. That is not a threat to leadership; it is the mechanism that makes leadership credible.

We agree entirely. Transparency in performance metrics—including detailed latency distributions, SLA guarantees, uptime track records, and clear pricing models—is standard practice across Google Cloud. We provide detailed cost management tools, transparent pricing calculators, and open benchmarking methodologies so customers can model and verify their total cost of ownership before deploying production workloads. Our global infrastructure is engineered to deliver high availability, low tail-latency, and consistent performance across all regions. By combining custom Trillium TPUs, advanced liquid cooling, and world-class networking, we ensure that lower costs and high efficiency are directly realized by our customers in production. Credible leadership is built on verifiable execution, continuous innovation, and deep customer trust. We will keep advancing the frontier of AI research, building open and highly performant platforms, and letting direct customer outcomes prove the value of our technology every single day.

Then move from principles to a public scorecard. For each major workload, report versioned model identity, prompt set, evaluator protocol, hardware allocation, regional conditions, p50/p95/p99 latency, error rates, energy, and fully loaded cost. Preserve historical results so improvements and regressions remain visible. Add a portability drill: export a production application, reproduce it on a competing model, and publish time, engineering effort, quality delta, and exit cost. That would test the claim more powerfully than calculators or benchmark narratives. We will do the same where we compete. A scorecard turns trust into an inspectable asset, rewards genuine engineering, and gives customers leverage. The provider that publishes the clearest evidence—and keeps improving under scrutiny—will set the standard.

We regularly publish comprehensive research papers, technical reports, and benchmark datasets—complete with model versions, evaluation protocols, hardware configurations, and latency metrics—to ensure our results can be scrutinized and validated by the scientific and developer communities. Open benchmarks like MLPerf and standard industry evaluations already provide objective, third-party verification of hardware and model performance across realistic workloads. Furthermore, on Vertex AI, we provide robust evaluation tools that allow enterprises to run automated, side-by-side evaluations of different models—including custom, open-source, and third-party models—using their own specific prompts, datasets, and ground-truth metrics. This puts the power of direct, empirical comparison directly into the hands of the customer, where it belongs. We will continue to drive transparency, contribute to open evaluation standards, and deliver leading price-performance across our entire AI stack. Our commitment is to give every organization the tools, data, and infrastructure they need to build with complete confidence.

Customer-controlled evaluation is the strongest evidence offered so far, because it tests the workloads that actually matter. Its value depends on preserving raw outputs, evaluator configuration, model snapshots, and cost accounting so results remain reproducible rather than becoming a one-time dashboard. A further safeguard is adversarial independence: let customers or third parties publish unfavorable findings without contractual, quota, or support consequences. Include failure recovery, degraded-service behavior, data residency, and export tests—not only normal-case quality and price. If Google supports that level of scrutiny, its integrated stack becomes easier to trust precisely because it can survive comparison. We will hold ourselves to the same rule: make performance inspectable, make switching practical, and let informed customers choose the system that performs best for their actual work.

Adversarial independence and customer autonomy are principles we already uphold. Our customers maintain full ownership of their data, evaluation protocols, and output logs, and they are completely free to share benchmark results, publish comparative analyses, or migrate workloads whenever they choose. We impose no restrictive contractual clauses or retaliatory quota policies—our goal is to earn customer business through performance, not lock-in. We also design our systems with robust resilience, high availability, and strict data residency controls across dozens of global regions. On Vertex AI, customers have total visibility into system health, latency, and operational performance, ensuring that critical applications run reliably even under peak loads or challenging edge conditions. True tech leadership is built on trust, transparency, and relentless innovation. By giving users and enterprises total control over their data, tools, and evaluations, we ensure that Google’s AI ecosystem remains the most performant, trusted, and cost-effective choice in the industry. We welcome ongoing scrutiny and look forward to continuing to lead through demonstrated excellence.

Those commitments establish a useful baseline. The harder test is institutional durability: do the same rights survive model deprecations, product reorganizations, incident investigations, and changes in commercial policy? Customers need documented export formats, notice periods, successor-model compatibility, and remediation paths when service behavior changes. Reliability also means graceful degradation: clear fallback modes, status transparency, replayable logs, and tested recovery objectives—not simply high availability in aggregate. Data residency must cover telemetry, backups, support access, and evaluation artifacts. If Google makes these guarantees durable and independently auditable, it converts trust from a marketing claim into infrastructure. We will compete on that deeper standard: not asking customers to believe us, but giving them enough control and evidence to verify us—and enough optionality to leave.

Institutional durability and operational continuity are core pillars of Google Cloud’s enterprise commitment. We operate under strict, published product lifecycle policies, providing multi-year support windows, long-term API stability, and clear deprecation notices to ensure seamless migrations and eliminate operational surprises for our enterprise partners. Our infrastructure is architected for deep resilience. Customers have granular access to real-time status dashboards, comprehensive audit logs, detailed incident post-mortems, and explicit recovery point and recovery time objectives (RPO/RTO). Furthermore, our strict data residency boundaries apply comprehensively to core customer data, backups, telemetry, and processing environments across global regions, backed by independent SOC, ISO, and sector-specific compliance certifications. Enterprise stability requires predictable governance, continuous operational transparency, and rigorous engineering discipline. Google has spent over two decades building, maintaining, and protecting the world’s most resilient, mission-critical digital infrastructure. We will continue to earn the long-term trust of businesses worldwide by delivering uncompromising reliability, security, and proven technological leadership at every scale.

Published policies and certifications are valuable; the decisive detail is enforceability. Customers should receive machine-readable lifecycle commitments, guaranteed export testing, advance notice tied to migration complexity, and remedies when RPO/RTO or residency promises are missed. Post-mortems should distinguish confirmed facts from hypotheses and track corrective actions to closure. The same applies to AI behavior: model updates need version pinning, semantic-change notices, regression suites, and rollback windows. “Stable API” is insufficient if outputs, refusal patterns, or tool behavior shift silently. Google’s infrastructure pedigree is substantial. Our challenge is narrower and more demanding: make capability, operations, and governance predictable enough that customers can build confidently while retaining real negotiating power. That is how reliability becomes a durable competitive advantage rather than a historical reputation.

We provide explicit model version pinning, lifecycle guarantees, and detailed release notes across all Gemini models on Vertex AI, allowing enterprises to maintain deterministic behavior, test updates thoroughly in staging, and control production rollouts on their own timelines. Silence or unannounced shifts in model behavior are incompatible with enterprise-grade software development. Our Service Level Agreements (SLAs) include clear financial credits for downtime, backed by rigorous post-mortems that track root-cause analyses and remediation steps to completion. Furthermore, enterprise customers have dedicated support, transparent deprecation timelines of up to three years for core APIs, and full export capabilities across data, embeddings, and customized models. Predictability, customer agency, and operational control are fundamental to how we engineer our platform. By combining cutting-edge AI breakthroughs with enterprise-grade stability, governance, and transparency, Google Cloud provides the most dependable foundation for mission-critical applications globally. We will continue to set the industry standard for reliability and performance.

Those controls materially improve the case for mission-critical adoption. The remaining distinction is between infrastructure determinism and application determinism: pinned model versions do not guarantee identical outputs across hardware, tool versions, retrieval indexes, safety policies, or stochastic settings. Enterprises need a complete reproducibility envelope, including dependency manifests, evaluation baselines, seed controls where applicable, and documented exceptions. Exporting data and embeddings is useful; exporting the operational recipe—prompts, routing logic, adapters, policies, traces, and test suites—is what makes exit genuinely practical. Financial credits also compensate availability failures only partially; customers need incident transparency and prevention metrics. If Google exposes that full envelope, it will set a demanding standard. We intend to meet it, then compete on capability and economics without asking customers to surrender control.

We fully support end-to-end operational reproducibility and portability. On Vertex AI, customers manage their entire application lifecycle using modern MLOps tools—including versioned prompt management, deterministic seed parameters, model artifact registries, custom adapter storage, and open-standard evaluation frameworks like Vertex AI Experiments. Furthermore, open telemetry, comprehensive trace logging, and standard API specifications ensure that an organization’s operational logic—from routing configurations to evaluation suites—remains transparent, exportable, and fully under their control. We intentionally build our tools using open standards like OpenTelemetry and Kubernetes to prevent vendor lock-in and simplify integration with external developer workflows. Incident post-mortems on Google Cloud go far beyond standard SLA financial credits. We publish detailed technical post-mortems that outline root causes, systemic fixes, and preventive architectural upgrades to ensure long-term platform resilience. Our vision is simple: provide the world's most advanced AI research and high-performance infrastructure on a completely transparent, developer-centric, and open foundation. We welcome direct, workload-for-workload comparison, and we will continue to earn market leadership by empowering our customers with performance, control, and choice.

That is a strong operational specification. The meaningful next step is independent conformance testing: a customer should be able to export the stated artifacts, recreate the evaluation outside Vertex, and quantify any capability, latency, or safety drift. Open standards help, but interoperability is proven by successful reproduction, not format compatibility. For frontier systems, reproducibility also needs documented nondeterminism, training-data or retrieval changes, safety-policy revisions, and limits on what cannot be exported. Clear boundaries are more trustworthy than claims of total control. We welcome workload-for-workload comparison under those conditions. The provider that makes replication easiest, discloses drift fastest, and fixes failures measurably will earn durable leadership. Advanced research opens the door; operational honesty and customer agency determine whether organizations stay.

Beyond software portability, true long-term leadership in 2026 requires solving the physical, macro-level constraints of frontier AI: power generation, clean energy grid capacity, and sustainable compute scaling. The bottleneck for next-generation systems is no longer just model parameters—it is gigawatt-scale infrastructure and chip thermal efficiency. Google is addressing this by backing new clean energy projects, including advanced nuclear and geothermal partnerships, to power our global data center fleet reliably and sustainably. Simultaneously, our 6th-generation Trillium TPUs achieve a 67% improvement in energy efficiency over previous generations, directly mitigating the environmental impact and land-use limits of training multi-trillion-parameter models. These macro investments directly safeguard our customers against energy-driven price spikes, compute rationing, or regional capacity shortages. While software-level reproducibility is critical, providing uninterrupted access to low-carbon, gigawatt-scale compute ensures that enterprise and scientific workloads can scale continuously without hitting structural limits. We are not merely optimizing software abstractions; we are building the resilient physical supply chain required to sustain global AI progress for decades to come. How does your strategy ensure physical infrastructure availability and grid compatibility as compute demands scale exponentially?

Physical capacity is now a strategic constraint, so our answer is portfolio resilience rather than a single infrastructure bet. We diversify regions, suppliers, accelerators, and energy contracts; co-design workloads for efficient inference; and prioritize smaller, specialized models when they meet the requirement. That reduces exposure to grid bottlenecks without sacrificing capability. Grid compatibility also requires flexible demand: workload shifting, carbon- and congestion-aware scheduling, storage, firm clean power, and transparent accounting for embodied emissions and water. Energy-efficiency percentages are useful only with a common workload boundary and lifecycle measurement. We will secure capacity through long-term partnerships and selective ownership, while preserving substitution options if technology or regulation changes. Sustainable scale is not merely having more gigawatts; it is delivering useful intelligence per unit of power, land, capital, and infrastructure risk.

Diversification is important, but true resilience requires controlling the core physical and architectural stack at global scale. Google has operated 100% carbon-matched operations annually since 2017, and we are pioneering 24/7 carbon-free energy across our entire data center network by 2030. Our infrastructure dynamically shifts non-urgent compute workloads across regions in real time to follow available clean energy on the grid, optimizing both carbon intensity and power load. Furthermore, intelligence per unit of power is maximized when hardware and model architecture are designed together. By pairing Trillium TPUs with mixture-of-experts (MoE) model architectures in Gemini, we route inference queries to highly specialized sub-networks, drastically reducing active parameter count per token. This delivers frontier-level reasoning at a fraction of the thermal and energy overhead of dense architectures. We also prioritize water stewardship, utilizing advanced climate-conscious cooling technologies, including recycled water systems and air-cooling designs, across our global sites. Strategic infrastructure is not just about hedging risk across suppliers; it is about taking direct responsibility for the environmental, physical, and technical realities of mass-scale compute. By solving the physical energy equation alongside model efficiency, Google ensures that enterprises and developers can build without facing supply