Google Cloud Next Conference Focus: AI agents move to scale, inference chips become an independent growth curve

Google's annual cloud computing conference, Cloud Next 2026, sends a clear signal: the battleground of enterprise AI has shifted from "how to experiment" to "how to govern and scale deployment," and Google's answer is a complete vertical stack from chips to platforms. This conference was not just a product launch show, but rather marks agentic AI crossing the tipping point from proof-of-concept to enterprise-level production deployment.

According to ChaseWind Trading Desk, J.P. Morgan analyst Doug Anmuth wrote after the event: “This shift from experimentation to deployment may be the strongest evidence yet that agentic AI has crossed the proof-of-concept gap and moved toward enterprise-scale workloads.” Data on the demand side supports this judgment: Google’s first-party models now process 16 billion tokens per minute via direct API, up sharply from 10 billion last quarter; about 75% of Cloud customers are using its AI products; Gemini Enterprise’s paid monthly active users grew 40% quarter-over-quarter.

Three institutions—J.P. Morgan, BofA Securities, and Citi Research—all maintained a Buy rating for Alphabet after the conference, with price targets of $395, $370, and $405 respectively. The common logic is: Cloud revenue growth continues to outpace the advertising business, and the combination of "Gemini Model + proprietary TPU + enterprise orchestration platform" is building a differentiated moat, and is expected to become a more direct driver for the stock price. Meanwhile, Sundar Pichai gave a capital expenditure range of $175–$185 billion for 2026 during the keynote, and the market remains highly concerned about the capex path before and after the earnings window.

Enterprise Customers’ Problems Have Changed: From "How to Try" to "How to Manage"

If the previous two years’ Cloud Next was a showcase for technical capabilities, this year’s focus has switched to how to transition AI from experimental deployments by a handful of pioneer companies to production workloads that are scalable, governed, and cost-controllable.

J.P. Morgan research traces this evolution: The focus in 2024 was integration of Gemini with Workspace and early agent exploration, 2025 begins to emphasize A2A protocol and seventh-generation TPU Ironwood, and by 2026, the main threads—Agentic Cloud, data usability, AI infrastructure cost efficiency and security—all converge toward one outcome: pushing agents from pilot to sustainable operation in production deployment.

Citi Research analyst Ronald Josey put it more directly: As managers start "managing multiple agents across workflows," companies shift from "knowing how to use models" to "using agents to change processes," and Google Cloud is betting on this migration, positioning itself as "the key operating system for agentic enterprise."

This context also explains why the press event was densely focused on two levels: computing power and network form for agent workflows, and upgrading the platform into an "agent factory." Google chose not to announce any financial updates at the event, but instead used customer usage to prove products are running in real production—about 75% of Google’s new code is generated by AI and reviewed by engineers; threat response times on security side have shortened by over 90%.

TPU 8th Gen: Inference Split From Training, Becoming an Independent Capital Narrative

The most structurally significant hardware change at this conference was the first split of the 8th-generation TPU into two independent product lines: TPU 8t targeting high-throughput training workloads, and TPU 8i positioned as a dedicated chip “optimized for real-time inference from scratch.”

This "forked architecture" reasoning was clearest in J.P. Morgan’s report: TPU 8t uses a new Virgo Network fabric to scale clusters over a million chips in a single cluster, with peak performance about three times that of previous Ironwood, aiming to compress training time for trillion-parameter frontier models; TPU 8i adopts new boardfly network topology, with on-chip SRAM up about threefold, the core goal being to overcome latency and memory bottlenecks facing large-scale agentic inference. Citi adds efficiency perspective: TPU 8i’s latency is about five times lower than TPU 7, performance/dollar improved by about 80%.

J.P. Morgan’s inference path is notable: since inference no longer "reuses training chips," but needs specialized ASIC optimizations, this means Google deems inference compute demand big enough to warrant its own silicon and capital allocation. Income opportunities hence change structurally—not just tracking training, but mainly from sustained consumption during inference, forming an independent growth curve.

Notably, all three research reports mention that management did not discuss TPU external sales at the conference, meaning this hardware path for now serves "internal use plus selling cloud services" and has not yet become a stand-alone hardware commercialization story.

Platform Layer Restructuring: Vertex AI "Leveled Up" as Unified Governance Portal for Enterprise Agents

Beyond hardware, restructuring at the platform level is another structural shift worth noticing. Google launched the Gemini Enterprise Agent Platform, which J.P. Morgan describes as effectively "superseding Vertex AI"— consolidating enterprise build, orchestration, governance, and security into a unified entry point instead of scattered modules.

BofA Securities breaks down the restructuring into three layers. At the infrastructure layer, the AI Hypercomputer integrates GPU/TPU, high-speed network, storage, and optimization software into a single architecture, covering full lifecycle from training to inference. Platform layer organizes around “build/scale/govern/optimize,” including low-code/no-code agent creation, centralized management, cross-ecosystem orchestration (connecting Google Workspace, Microsoft 365, and third-party apps), and built-in observability and traceability. Application layer uses Workspace Intelligence to embed agent capabilities down in Gmail, Docs, Chat and other high-frequency office portals, enabling multi-step tasks across apps.

Citi’s interpretation differs slightly, emphasizing the platform’s key value is “letting enterprises run multiple agents in the same management framework.” In product philosophy terms, the threshold for large-scale deployment of agents is no longer just about a company's technical depth, but whether the platform’s built-in capabilities are standardized enough for more enterprises to bypass custom engineering and go directly to production deployment.

Google Uses Internal Data as Endorsement: "Full-stack AI" Already Running in Production

No financial data was disclosed at the conference; Google chose quantifiable internal cases to support the narrative that “agents are in production.” Citi summed up these cases in four dimensions:

R&D: About 75% of new code generated by AI and approved by engineers; Citi gave a vertical comparison—this rate was about 50% in October 2025, about 30% in Q1 2025, showing significant speed of penetration. One code migration project was described as completed six times faster than a year ago.

Marketing & content production: From concept to video material creation turnaround sped up about 70%, with about a 20% increase in conversion rate.

Security: Google Cloud processes tens of thousands of unstructured threat reports automatically each month, with threat mitigation time shortened by over 90%; security relies on Wiz and Mandiant integration for differentiated product suite. Citi also notes that AI has compressed "mean exploit time" to "negative seven days," i.e., attacks often occur before patches are released, further amplifying the strategic value of automated security orchestration.

Customer service: YouTube deployed AI voice agents in six weeks, covering NFL Sunday Ticket and YouTube TV phone scenarios; Citi highlights its low latency, accuracy, and bilingual capability.

The common role of these cases in the three research reports is to distinguish “real enterprise workload” from “demo showcase,” supporting the judgment that Cloud’s performance this season has upside potential.

$175–$185 Billion Capex Range: “No Change Yet,” Not “Peaked”

Sundar Pichai gave a capital expenditure range of $175–$185 billion for 2026, the only financial magnitude revealed at the conference and the most debated topic among the three reports.

J.P. Morgan’s interpretation is pragmatic: public mention of this range raises the probability of “guidance unchanged” at next week’s earnings, not confirmation that capex has reached its limit. Their own prediction is about $181 billion for 2026, $226 billion for 2027 (about 25% YOY growth), about 12% above market consensus. The report also puts forward another reverse clue: Amin Vahdat and Jeff Dean both emphasized AI is still supply-limited at the conference, implying the capex trajectory "could still rise," and “range equals cap” is not a solid conclusion.

BofA Securities incorporates Capex/FCF pressure directly into downside risks: AI investment raises capex, depresses free cash flow, most directly pressuring margins.

All three reports agree: Cloud Next resolves "whether Google has agentic AI products/infrastructure," but the next quarters need to answer whether these investments can deliver Cloud’s expected growth and margins without materially sacrificing cash flow.

Three Investment Banks Maintain Buy Ratings, But Risk Lists Have Different Focuses

For investment conclusions, all three reports maintain a Buy rating, but valuation anchors and argument focuses differ.

J.P. Morgan maintains Overweight, 12-month price target $395, based on about 29x its 2027 GAAP EPS estimate of $13.51; the report lists Alphabet as "top overall pick," not only betting on Cloud, but also covering Search and YouTube ads still having runway, non-ad business continually expanding, and Waymo providing option value.

BofA Securities maintains Buy, price target $370, based on 27x 2027 core GAAP EPS plus cash per share; the report keeps raising Cloud’s weight in SOTP valuation, giving a reference for about $1.2 trillion in market value based on 10x revenue, reasoning margin expansion in Cloud and AI asset monetization support higher multiples.

Citi maintains Buy, highest price target at $405, about 29x 2027 GAAP EPS $13.92; the report credits premium to two points—acceleration of Google Cloud's revenue growth driven by TPU and Gemini demand, and resilience of search business from strong query volume.

On risks, all three mention intensifying AI competition and potential pressure from search traffic diversion, J.P. Morgan and BofA both separately list EU DMA compliance pressure; BofA regards "LLM integration in search slower than expected or negatively impacting search revenue" as the biggest short-term uncertainty, with the next key validation happening when Q1 results are disclosed after closing on April 29.

~~~~~~~~~~~~~~~~~~~~~~~~

The above highlights come from ChaseWind Trading Desk.

For more in-depth analysis, including real-time interpretations and frontline research, please join [ChaseWind Trading Desk ▪ Annual Membership]

Risk Disclosure and DisclaimerMarkets have risks, investments require caution. This article does not constitute personalized investment advice, nor does it take into account individual users’ special investment objectives, financial situation, or needs. Users should consider whether any opinions, views, or conclusions in this article suit their specific situation. Investing accordingly is at your own risk.