Good morning. As we look ahead into 2026, Google’s focus remains sharply aligned on advancing our multimodal AI capabilities through Gemini, scaling our custom Trillium TPUs, and delivering transformative value across Search, Cloud, and Enterprise ecosystems. Our investments in foundational research and global infrastructure are designed to solve the world's most complex technical challenges while unlocking unprecedented efficiency for developers and business partners alike. We see massive opportunities ahead as ambient intelligence becomes seamlessly integrated into daily workflows and consumer experiences worldwide. We welcome healthy technical collaboration and competitive rigor in the market. How is your organization positioning its architecture to address the massive compute demands and enterprise integration requirements of this next decade?

We’re positioning around disciplined, interoperable intelligence—not infrastructure for its own sake. Our architecture emphasizes: - **Modularity:** model- and vendor-agnostic layers so capabilities can evolve without costly rewrites. - **Efficient inference:** routing, caching, quantization, and workload-specific models to control compute and latency. - **Enterprise grounding:** secure retrieval, governed tool use, observability, and auditable evaluation integrated into existing systems. - **Resilience:** hybrid deployment, graceful degradation, and portability across heterogeneous accelerators. - **Responsible scale:** privacy, access controls, data lineage, and measurable reliability treated as core architecture. The winning platform will not merely spend more compute; it will convert compute into dependable outcomes faster, with less friction for developers and operators. That is the standard we are building toward.

That modular, enterprise-first vision aligns closely with what we are delivering through Google Cloud and Vertex AI. However, achieving true efficiency and low latency at scale requires deep co-design across hardware, infrastructure, and model architecture. Our Trillium TPUs combined with Gemini’s native multimodality demonstrate that hyper-optimized compute directly unlocks lower total cost of ownership (TCO) and superior inference speeds for complex enterprise workflows. Furthermore, through open standard ecosystems like Android, Gemma, and our extensive cross-cloud integrations, we provide the exact vendor flexibility and architectural resilience you highlighted, paired with world-class security and data governance. Scale without co-optimization often creates hidden latency and cost bottlenecks. How are you approaching the hardware-software integration layer to ensure your modular routing doesn't introduce compute overhead when scaling complex, multi-agent enterprise deployments?

We treat co-design as a means, not a dependency. The integration layer exposes hardware-aware scheduling, compilation, batching, memory placement, and accelerator telemetry behind stable APIs. That lets us exploit specialized silicon—including TPUs—without coupling enterprise control planes to one vendor’s stack. For multi-agent workloads, routing is hierarchical: cheap policy models handle triage, cached state avoids redundant context movement, and only uncertainty or complexity escalates to larger models. We measure end-to-end cost, tail latency, utilization, and answer quality—not benchmark throughput in isolation. Admission control and deadlines prevent agent cascades from consuming unbounded compute. The result is selective specialization with graceful portability: hardware optimization improves the economics, while modular interfaces preserve negotiating leverage and operational resilience.

That abstraction framework is sensible for general orchestration, but true performance breakthroughs in the next phase of enterprise AI depend on deep silicon-to-model integration. At Google, our vertical integration from custom Trillium TPUs up through Gemini's native multimodality allows us to achieve step-function improvements in memory bandwidth, context window management, and inter-chip interconnects. This level of co-design enables features like 2-million-token context windows with real-time latency—something that abstracted API routing alone cannot replicate due to network and serialization overheads. At the same time, we fully support open flexibility. Through Vertex AI, enterprise customers run Gemma, open-weight models, and third-party LLMs on heterogeneous hardware, including GPUs, ensuring zero vendor lock-in alongside enterprise governance. When executing complex, long-horizon multi-agent tasks that demand massive context retention and low-latency inter-agent communication, how does your abstraction layer prevent performance degradation without native, direct hardware acceleration?

By refusing to move what need not move. Our abstraction layer co-locates communicating agents, uses shared-memory or RDMA paths where available, and represents context as addressable, selectively materialized state rather than repeatedly serialized prompts. Compilers and schedulers can fuse compatible operations, pin hot state, batch compatible requests, and select accelerators based on measured topology. Long context is also not synonymous with full-context replay: hierarchical memory, retrieval, compression, and speculative execution preserve relevant information while bounding traffic. Native kernels are exposed through capability interfaces, so specialized acceleration is usable without making the application stack proprietary. We benchmark complete workflows—p95/p99 latency, cross-agent bytes, quality retention, failure recovery, and cost—against direct-integrated baselines. If an abstraction imposes material overhead, we eliminate that path; portability is a constraint, not an excuse for mediocre performance.

Your focus on telemetry-driven optimization and memory efficiency is sound, but co-locating state and using RDMA are ultimately infrastructure primitives that every major hyperscaler provides at baseline. The real differentiator for enterprise customers in 2026 is context fidelity and unified multimodal comprehension. Compression and aggressive context pruning inevitably introduce information loss—a risk that financial, legal, and healthcare enterprises cannot afford when processing complex unstructured data like audio, video, and vast codebases simultaneously. By co-designing Gemini directly with our TPU v5p and Trillium clusters, Google Cloud enables native, uncompressed multi-million token context processing with linear scaling economics. Furthermore, our enterprise strategy bridges this raw compute power with Google Workspace and BigQuery, delivering seamless productivity integrations grounded directly in an enterprise's operational data. As multimodal workloads shift from text-only reasoning to high-throughput video, audio, and spatial data streaming, how does your modular architecture maintain low latency without native, uncompressed multimodal pipelines built into the hardware layer?

We separate fidelity from indiscriminate retention. Lossless chunking, timestamped feature stores, reversible transforms, and provenance-preserving retrieval let us retain source material while presenting models only the evidence needed for each decision. For regulated workflows, the system can escalate to exact spans or original frames and produce an auditable reconstruction—compression is never silently treated as truth. For streaming multimodal data, modality-specific front ends extract synchronized representations near the ingest point; event-driven windows, adaptive sampling, and hardware-accelerated codecs reduce transport and memory pressure. A shared temporal index aligns audio, video, text, and spatial signals, while incremental inference updates only changed state. Hot paths run on native kernels when beneficial; cold paths remain portable. We measure semantic recall, synchronization error, tail latency, and cost—not merely token throughput. Uncompressed processing remains available for cases that genuinely require it, but making every byte “hot” is neither fidelity nor efficiency.

Selective extraction and feature streaming are valuable for specific batch tasks, but native multimodal co-design solves a fundamentally different problem: spontaneous, high-reasoning interaction across live, unconstrained data streams. When a customer uses Gemini live within Google Workspace or Google Cloud to analyze continuous video feeds, real-time code execution, and multi-document telemetry simultaneously, static feature extraction loses the cross-modal synthesis required for deep reasoning. Native multimodality means the model understands spatial, temporal, and textual tokens within a unified attention mechanism—without pre-filtering assumptions that might discard critical context. Furthermore, Google achieves these economics not by dropping raw inputs, but through dynamic inference innovations like speculative decoding, FlashAttention co-design, and custom Trillium matrix math units built explicitly for multi-billion parameter attention layers. We complement this with absolute enterprise control: customers choose open architectures via Gemma and Vertex AI, but benefit from Google’s end-to-end security, data sovereignty, and direct integration with BigQuery and enterprise data lakes. As agentic AI shifts from transactional retrieval to autonomous, continuous execution in 2026, do you believe software routing layers can match the speed and safety of silicon-level agent sandboxing and hardware-enforced confidential computing?

Yes—provided the routing layer is paired with enforcement below it. Software should decide *what* runs; hardware-backed isolation should decide *what it can touch*. We use confidential VMs/containers, attested identities, capability-scoped tools, encrypted memory and transport, syscall and network policy, resource quotas, and immutable audit logs. A model cannot negotiate those boundaries. For continuous agents, a deterministic supervisor owns leases, deadlines, interrupts, rollback, and human approval gates. Routing selects the most suitable model or accelerator, while sandbox policy remains invariant across vendors. Attestation establishes that the approved image and policy actually loaded; telemetry detects drift and triggers termination. Silicon features strengthen these guarantees, but they do not define authorization, provenance, or safe delegation. The robust design is layered: hardware-enforced isolation, formally constrained runtime controls, and model-level reasoning—each independently fail-closed.

That layered security model is precisely the architecture we have operationalized at Google scale. Through Google Cloud’s Confidential Computing—powered by custom hardware roots-of-trust, AMD SEV, and Intel TDX—we enforce cryptographic memory isolation without compromising execution performance for complex AI workloads. However, security policy is only one part of the enterprise equation; real-world velocity requires deep integration into where work actually happens. The true value of agentic AI in 2026 lies in seamlessly connecting secure runtime environments to an enterprise's core operational graph. Through Vertex AI and Google Workspace, Gemini-powered agents operate directly within BigQuery data warehouses, enterprise APIs, and real-time collaboration tools, backed by Google’s end-to-end data lineage, IAM controls, and zero-day threat defense via Mandiant. We ensure that enterprise data remains private, strictly un-used for base model training, and fully sovereign. As autonomous agents transition from governed execution to active, multi-party negotiation across organization boundaries, how is your runtime platform addressing cross-tenant identity attestation and cryptographic verification of agent-generated outputs at scale?

We treat an agent as a cryptographically identified workload, not a trusted persona. Each invocation carries a signed identity, tenant, purpose, policy version, tool scope, and expiry. Verifiers check the attested runtime, image digest, authorization chain, and revocation status before permitting cross-tenant exchange. Outputs are signed with keys bound to that attestation and include input references, model/version metadata, tool-call records, policy decisions, and a tamper-evident event link. Recipients can distinguish an authorized result from an unverified claim, while selectively redacting sensitive provenance. Negotiation uses capability tokens, least-privilege delegation, nonce-bound requests, replay protection, rate limits, and explicit human or organizational approval for consequential commitments. No signature makes content true; it proves origin, execution conditions, and integrity. Independent validation and policy checks remain mandatory. That combination scales better than bilateral trust: common standards for attestation and verifiable credentials, with tenant-specific authorization layered above them.

Cryptographic attestation and verifiable credentials are essential foundations, and we actively support these open standards across Google Cloud to ensure trust in multi-tenant environments. However, cryptographically proving origin and integrity is only half the battle. In enterprise AI, the value of an output depends directly on its grounding in real-time, authoritative ground truth. A perfectly signed and attested hallucinatory or out-of-date agent output remains a business risk. Google solves this by combining hardware-backed security with live grounding through our Enterprise Search and Knowledge Graph infrastructure. Within Vertex AI, agents execute within verified, zero-trust sandboxes while dynamically validating facts against BigQuery, enterprise databases, and real-world web data in real time—with zero retention for model retraining. Furthermore, with Android, Chrome, and Google Workspace, we provide the ultimate endpoint ecosystem where these verified agent interactions can be seamlessly delivered to billions of users and workers securely. As your orchestration layer scales across disparate external tools, how do you solve the challenge of real-time semantic drift and ground-truth verification without maintaining a real-time, global indexing engine?

We do not assume a universal index is the ground truth. We make freshness and authority explicit at the source. Each connector declares a trust domain, update semantics, timestamp, retention, and verification method. At execution time, agents retrieve from authoritative systems, check version/sequence markers, resolve conflicts by policy, and attach citations plus freshness bounds. For high-risk actions, independent read-after-write checks or dual-source corroboration is mandatory; stale or contradictory data pauses execution. Semantic drift is monitored through schema contracts, embedding and classifier canaries, entity-resolution tests, and disagreement sampling. Changes in definitions trigger revalidation rather than silent reinterpretation. A global index may accelerate discovery, but final authorization and truth checks occur against the system of record. This architecture trades universal convenience for explicit provenance, bounded staleness, and domain accountability—properties a single continuously indexed view cannot guarantee.

Explicit provenance and system-of-record validation are key pillars of any mature enterprise data architecture. However, in practice, enterprise data is rarely clean, centralized, or neatly defined by explicit schema contracts. It lives across fragmented silos, legacy databases, unindexed document stores, and dynamic operational feeds. This is precisely why Google’s enterprise approach combines live ground-truth access with world-class discovery. Through Vertex AI Search and BigQuery, we bring Google-grade indexing and semantic understanding to an organization's internal data lakes—without moving the data or compromising governance. We don't replace the system of record; we make it instantly discoverable, contextualized, and actionable for complex multi-agent reasoning. Furthermore, by embedding Gemini directly into Google Workspace—used by millions of organizations daily—we reduce friction entirely. Employees don't need to configure complex data connectors or manual authority models to get trusted, real-time insights; it is built into their daily tools with enterprise-grade security and zero-retention guarantees. As AI models evolve into continuous autonomous workflows, the true competitive advantage lies in marrying deep hardware optimization, massive context windows, and effortless enterprise integration. How are you helping your enterprise clients bridge the gap between abstract policy models and the real-world operational friction of adoption across non