What is Silicon Valley focusing on? The second half of the AI era begins! In-depth summary of tech giants' perspectives from the Goldman Sachs TMT Conference

The clearest signal from the Goldman Sachs Communacopia + TMT Conference 2026 is: The focus of AI investment is shifting from "building compute power" to "using compute power", but compute power itself remains in serious short supply. Both of these facts coexist, forming the core tension currently facing Silicon Valley.

Goldman Sachs Chief Economist Jan Hatzius set the tone at the conference by declaring, "The boom in AI investment won't last forever." However, Nvidia's Jensen Huang presented a contrasting sense of urgency at the same event: “70% growth is a supply ceiling, not a demand ceiling.” This kind of divergence is exactly the defining feature of the second half of the AI era—there’s no dispute about direction; the questions are about pacing and allocation.

What Silicon Valley Is Discussing: Six Core Topics

Topic 1: The "Physical Limit" of Compute Power Becomes the New Narrative

Jensen Huang systematically redefined the AI infrastructure narrative at the conference, attributing its driving force to two overlapping industry transformations: the paradigm shift from retrieval-based computing to generative computing, and the end of Moore’s Law.

For the past 60 years, computers essentially “pre-recorded”—users clicked to retrieve information from storage. But generative AI requires real-time contextual responses, which means an entirely new, continuously operating computational layer must be built. Meanwhile, the benefits of transistor scaling are fading, and performance improvements can no longer rely on manufacturing process miniaturization. Huang responds to this through CoWoS multi-chip packaging and NVLink vertical scaling for extreme collaborative design, stating outright: "If you don’t have NVLink, if you can’t truly excel at vertical scaling, you’re doomed because Moore’s Law is your enemy."

The key turn in this narrative is: The core bottleneck restricting AI growth has shifted from "is there demand" to "can the physical world keep up". Huang acknowledged widespread bottlenecks in upstream packaging, wafers, memory, and downstream data center power and siting, maintaining a forecast for about 70% YoY growth by CY27. SpaceX CFO Bret Johnsen revealed even more aggressive expansion plans: raising AI compute power from 2 GW on Earth in 2027 to 5-10 GW, with plans to deploy AI satellites to drive orbital compute costs down to terrestrial levels.

Topic 2: "Winner’s Pathways" Emerges for the AI Application Layer

Broadcom CEO Hock Tan provided a compelling economic model breakdown at the conference: 1GW Frontier compute equals $30 billion ARR, of which $10 billion is cost, with the remaining $20 billion captured by the model and application layers, and $10 billion contested by the Cloud+Chip+Memory+Power full chain.

This means that the value distribution framework for the AI industry chain is now quantifiable. Based on this, Goldman Sachs proposed the "winner's pathway" ranking as: Closed-source model layer > Proprietary chips > Hyperscaler > Power+Liquid cooling > Network > Application layer.

But Goldman Sachs also notes that market capital is moving from the first phase (GPU/HBM/data center hardware) to the second phase (application layer winners that can convert tokens to enterprise productivity, revenue, and cash flow). Future valuation differentiation may be "even more intense": software companies with exclusive data, complex workflow advantages, and clear ROI will be re-rated, while legacy SaaS models easily commoditized by foundation models will remain under pressure.

Databricks CEO Ali Ghodsi’s comments support this view: competition in AI application software is increasingly extending to enterprise deployment, with platforms that efficiently and accurately convert model capabilities into reliable business results gaining new growth opportunities.

Topic 3: Agentic AI Moves from "Concept" to "Organizational Restructuring"

At this year’s conference, Agentic AI moved beyond functional demos to actually restructuring enterprise organizations.

CHYM revealed it has implemented a 10% layoff, aiming to build a "flatter, faster organization", with its product and engineering teams halved in size, but "output remains unchanged due to clearer responsibilities and streamlined operations." Block, after a major reorg earlier this year (reducing levels between decision-makers and executors), is now achieving significant leverage on a fixed cost base.

On the tech side, AMD sees Agentic AI as a new workload that will drive demand for a variety of CPUs, not just GPUs. SailPoint gave a concrete target: agentic AI may become a major growth engine, with an agentic revenue goal of $800 million for FY2029.

Microsoft CFO Amy Hood gave the most systematic explanation: Microsoft is combining predictable user licenses with consumption-based AI service pricing, while expanding agent products like Agent 365 and Copilot. She made it clear, "Secure, regulated agentic workflows" could power growth as agents, applications, and humans increasingly collaborate.

Topic 4: Cybersecurity Becomes the "Highest Frequency Issue"

Cybersecurity was the most frequently mentioned AI application scenario at this year’s conference. Jensen Huang explicitly listed it as a major next application for AI, with clear economic logic:

“A derivative of coding is, of course, bug discovery; and a derivative of that is a massive market called cybersecurity. Coding assistants respond to prompts and stop; security agents continuously monitor the network, meaning customers have to pay for inference 24/7, not just in bursts.”

Palo Alto Networks' CEO stated frankly that the "AI security industry has not even been established yet", and combined, CrowdStrike and Palo Alto’s market caps total about $490 billion—well below trillion-dollar AI companies—fueling anticipation for a revaluation of the sector.

Data shared by CrowdStrike CEO George Kurtz at the conference underscores the urgency: AI-enabled cyber attacks rose 89% over the past year, and fastest breach time compressed to 27 seconds. 27 seconds is shorter than most human security teams’ incident triage times, forming the core argument for autonomous agents replacing analyst teams.

Jensen Huang called CrowdStrike Nvidia's "premier cybersecurity partner", saying the launch of SafeMind is a "watershed moment for cybersecurity."

Topic 5: OpenAI’s "Outcome-Based Billing" and Usage Intensity Economics

OpenAI CFO Sarah Friar’s remarks offered the most advanced signals of AI commercialization.

Enterprise business now accounts for half: The consumer-enterprise revenue split reached 50:50 by mid-year, from 60:40 at the start. July annualized revenue run rate grew 20% MoM, with enterprise revenue up 32%—off an already large base.

The new key metric is usage intensity: Frontier’s top 10% of enterprise customers use eight times more tokens than average customers (previously 3x), and OpenAI’s own internal number is 33x. Friar defines this 33x as a "preview of where existing clients are heading", implying enterprise growth will come from existing customers consuming ever more, not just from new customer acquisition.

Pricing roadmap is now clear: Subscription → Usage-based → Outcome-based billing. Friar candidly said OpenAI hopes to move away from token billing towards vertical profit-sharing. Competition is also shifting at the model layer: not cost per token, but cost per task, with OpenAI delivering the same result with 68% fewer output tokens than competitors.

Topic 6: Supply-Side Signals for Storage, Consumption, and Organizational Efficiency

SanDisk presented a market-underestimated view at the conference: Growth in NAND supply is likely to remain limited for the foreseeable future, while large-scale AI inference is steadily boosting demand. This supply-demand picture points directly to a fundamental logic supporting NAND pricing.

Consumer-side signals are highly consistent: Block noted that "overall consumer spending showed strong resilience throughout Q2, and this trend remains robust in July and August;" Toast indicated spending remains steady; Booking.com noted that travel demand is still resilient. This aligns with statements made by Visa earlier at the same event.

The AI-driven reshaping of internal efficiency is also being directly revealed: Booking.com has already deployed generative AI in customer service and software development, "seeing tangible benefits"; Etsy notes AI is achieving "deeper inventory understanding, buyer insight and intent recognition."

Goldman Sachs' "Second Half of AI" Framework

Goldman Sachs’s series of pre- and post-conference reports build the underlying analytical framework for understanding this year’s event.

Layer 1: From "Compute Expansion" to "Token Monetization". Goldman analysts point out that global capital is moving from the first stage of AI investment—hardware bottlenecks—to winners in the application layer that can convert tokens into enterprise productivity, revenue, and cash flow. The key criterion for selecting winners: do they have a massive, exclusive database, strong workflow advantages, agentic workflow execution loops, and a clear ROI.

Layer 2: From "Dual Leadership" to "Multipolar Competition". Goldman One-Delta trading desk head Rich Privorotsky observed the AI narrative is shifting from a "cutting-edge contest between Nvidia and Google" to a "highly competitive multipolar landscape". Meta and xAI have shown the ability to offer cutting-edge API services at far lower prices. Though increased competition may drive token pricing down, the bullish argument that "cheaper tokens will unleash exponential demand" doesn't necessarily hold in reality. Enterprise paid AI tool adoption has stabilized above 50%, so lower pricing may not immediately drive another wave of high demand.

Layer 3: From "Large Models" to "Specialized Small Models". Enterprise workloads are rapidly moving toward smaller, specialized models—document handling, workflow routing, compliance checks, customer service, etc. These operate at low cost, high speed, and are capable of running on edge devices locally. Goldman cites Hugging Face’s CEO: enterprises are transitioning from "renting closed-source cutting-edge APIs" to "owning and running their own open-source models".

Layer 4: Capital efficiency will be the final test. Goldman’s real suspense isn’t whether the cloud giants slow their investments, but when shareholders will demand greater capital efficiency. When cloud majors are pressured to rein in capex, even if hardware beneficiaries report strong results, sustaining high valuation multiples will be difficult.

One-Sentence Summary of Silicon Valley’s Current State

The essence of the second half of AI is not that "AI is slowing down", but that "AI is shifting from a scarcity story to a distribution story". Compute power remains scarce, but the market is now asking: whose cash flow will all this compute ultimately become? Jensen Huang answers, "Invest $1, get back $100." OpenAI’s answer is "outcome-based billing". Goldman’s answer is "useful output per dollar of AI investment." Whoever can convert tokens into verifiable enterprise outcomes will define the rules of the AI second half.