From Training to Inference: NVIDIA's Earnings Ignite a Paradigm Shift in AI Computing

This week, the global AI industry witnessed what can only be described as a true "super week." The catalyst was NVIDIA's quarterly earnings report, released after the U.S. market closed on August 26—a document destined to be recorded in semiconductor history. It not only sent shockwaves through financial markets but also, with irrefutable data, revealed an accelerating structural transformation: the primary battlefield of AI compute is shifting decisively from large‑scale model training to mass‑scale inference deployment. This is not merely a technical realignment; it is a fundamental redefinition of the industry's business models and growth logic.

A "Crushing" Earnings Report in Full View

Let us begin with the core financials. For the second quarter of fiscal year 2027 (corresponding to the natural calendar second quarter of 2026), NVIDIA posted total revenue of **$96.221 billion**, representing a year‑over‑year increase of 106% and a sequential rise of 18%. This figure far exceeded the upper end of consensus estimates, which had been hovering around $92 billion. Within this, data center revenue reached $89.0 billion, up 117% year over year and accounting for 92.5% of total revenue, further cementing its role as the company's absolute growth engine. On the bottom line, GAAP net income came in at $59.688 billion, a 126% surge from the prior year, while diluted earnings per share also hit an all‑time high.

Every single metric surpassed even the most bullish sell‑side projections. Yet what truly elevated market sentiment from "satisfied" to "euphoric" was NVIDIA's remarkably rare long‑term guidance. The company projected that revenue for the full fiscal year 2028 would grow by approximately 70%, whereas the average market estimate had been only around 45%. Such an extended, two‑fiscal‑year optimistic outlook is exceptionally rare for NVIDIA, and it sent a clear signal: management possesses extraordinary visibility and confidence in compute demand for at least the next 18 months.

Equally noteworthy is the pace of hardware generational transition. The much‑anticipated Vera Rubin platform, hailed as the "next‑generation AI computing foundation," entered full‑scale production and shipments in August. During the earnings call, NVIDIA's CFO revealed that the platform is expected to contribute roughly 20% of data center revenue in the third quarter (i.e., August–October 2026). Multiple Wall Street analysts pointed out that the third quarter of 2026 will mark a critical industry inflection point for global AI hardware, signaling that the entire supply chain is officially moving from the previous Hopper‑centric cycle into a new era of mass‑production execution centered on Vera Rubin. Following the release, NVIDIA's stock surged 8.74% in a single day, driving AMD, Broadcom, Micron Technology, and other AI hardware stocks broadly higher, with the Philadelphia Semiconductor Index climbing more than 3% in tandem.

Inference: The Redefined "Growth Engine"

If the outstanding numbers were the surface, then Jensen Huang's characterization during the earnings call was the soul. He unequivocally stated: "Inference is the growth engine of AI." This was no mere placating slogan; it was grounded in a profound understanding of the explosive evolution occurring at the application layer.

Over the past three years, AI investment had been overwhelmingly skewed toward training. Tech giants and startups alike rushed to acquire tens of thousands of GPU clusters to pre‑train large language models with hundreds of billions or even trillions of parameters. However, training is, by nature, a one‑time capital expenditure—once a model is trained, the marginal cost falls to near zero. Inference is fundamentally different. Every user interaction with ChatGPT, every automated task executed by an AI agent, every personalized summary generated by a search engine, and every frame of AI‑generated video continuously consumes inference compute. This consumption is perpetual, cumulative, and grows exponentially with rising user activity and application penetration.

As AI agents, intelligent search, code assistants, real‑time translation, and other applications begin to scale, the growth rate of inference demand is now numerically surpassing that of training. NVIDIA has responded with a highly targeted hardware weapon—the Groq 3 LPX inference chip, which has entered full production. This inference‑specific accelerator, deeply customized on the Vera Rubin architecture, achieves a processing rate of 3,400 tokens per second when running the latest Gemma 4 31B open‑source model—four times faster than current mainstream alternative platforms. Its core design philosophy is no longer about maximizing general‑purpose matrix multiplication peak performance, but about delivering extreme performance for agentic AI workflows and latency‑sensitive real‑time interactions. This marks a strategic shift: NVIDIA is no longer content to cover all compute scenarios with a single universal GPU; it is beginning to follow the playbook of CPU giants by tailoring chips for specific workloads. Nebius, an AI cloud platform that is among the first to adopt this chip, saw its stock rise more than 6% after hours—the market is voting with its wallet for the "inference‑specialized" segment.

Mega‑Deals and the Global Infrastructure Race

NVIDIA's stellar earnings were merely the tip of the iceberg in this week's frenzy of AI compute activity. In the days surrounding the report, a flurry of astronomical compute‑locking deals emerged, revealing an almost frantic scramble for resources among industrial capital.

Most eye‑catching was the announcement by AI unicorn Anthropic that it would invest $45 billion to lease computing capacity from cloud infrastructure provider Nscale at its data center campus in West Virginia. The six‑year agreement corresponds to approximately 460 megawatts of power capacity, and Nscale will deploy NVIDIA's latest Vera Rubin chips across the cluster. This transaction not only sets a record for a single AI startup's compute lease but also underscores that even leading large‑language‑model developers heavily depend on external, specialized compute operators for leverage.

Meanwhile, cloud titan AWS and NVIDIA announced a significant expansion of their strategic partnership. The two companies plan to deploy an additional 2 million NVIDIA GPUs across AWS's global infrastructure between 2027 and 2028. This includes integrating NVIDIA's next‑generation Vera CPU into AWS services and building a massive AI compute factory containing 100,000 GPUs for the U.S. government's dedicated cloud environment. Such deep‑rooted co‑development between cloud providers and chip designers is reshaping the entire supply chain upstream of cloud computing.

Turning to China, the pace of infrastructure build‑out is also accelerating notably. According to planning information revealed by national authorities, the "15th Five‑Year Plan" period will see an additional direct investment of RMB 4 trillion in computing network construction, with a focus on compute‑power synergy, ultra‑large intelligent computing cluster interconnection, and other frontier areas. By the end of June this year, China's total intelligent computing scale had reached 2,185 EFLOPS, and more than 70 cross‑regional compute corridors had been established around the eight national hub nodes. In the Yangtze River Delta region, a large‑scale green intelligent computing center with a total investment of RMB 7 billion, designed to house a 100,000‑card class cluster, is steadily advancing through civil works. Additionally, the Hohhot Green Computing Summit saw on‑site contract signings totaling RMB 136.1 billion; in the first seven months of this year, more than 150 billion‑yuan‑level intelligent computing projects were signed or commenced nationwide. Together, these figures outline a three‑dimensional computing infrastructure network spanning national top‑level design, local industrial implementation, eastern application‑dense regions, and western energy‑rich areas.

Industrial Performance: From Thematic Speculation to Tangible Delivery

NVIDIA's high‑quality earnings report validates a pivotal judgment for AI investment—not only in China's A‑share market but globally: the development logic of the AI compute industry has completely transitioned from the policy‑expectation and concept‑driven phase of 2025 to early 2026, into an era of substantive performance delivery backed by orders and cash flows.

Mid‑year financial reports from Chinese A‑share listed companies offer compelling evidence. Industrial Fortune Union, a leading AI server manufacturer, posted first‑half revenue of RMB 557.861 billion, a year‑over‑year increase of 54.63%, with net profit growth accelerating in tandem. Inspur Information expects first‑half net profit to grow between 226% and 288% year over year. Eoptolink, a front‑runner in optical transceivers, reported a 241.70% increase in net profit for the first half, with gross margins continuing to improve thanks to the ramp‑up of 800G and 1.6T high‑speed optical modules. Aggregating across sub‑sectors, the total net profit of A‑share digital chip design companies for the first half surged by a staggering 405.64% year over year. These are no longer distant, intangible concepts; they are concrete figures reflected in cash flow statements and income statements.

The Paradigm Shift Redefines the Compute "Ceiling"

The formal arrival of the inference era means that the "ceiling" of the overall AI compute market will be redefined and substantially elevated. Training demand is essentially a lumpy capital expenditure—its growth curve is constrained by the pace at which large customers build data center clusters; once infrastructure is in place, procurement naturally tapers off. Inference demand, in contrast, is a classic operating expenditure with strong persistence and stickiness—as long as global monthly active users of AI applications continue to grow, and as long as more AI agents are embedded into every enterprise's digital processes, inference consumption will only monotonically increase, with virtually no cyclical downturn.

This, perhaps, is the underlying mathematical logic behind NVIDIA's bold two‑year, 70% revenue growth guidance. Although this week's "AI Super Week" has drawn to a close, the profound implications of the training‑to‑inference paradigm shift revealed by NVIDIA's earnings are only beginning to permeate every capillary of the industry chain. Future compute competition will no longer be solely about who possesses the most H100 or Vera Rubin chips; it will be about who can deliver the lowest marginal cost, the lowest latency, and the highest energy efficiency to serve billions of end‑users in their unconscious, everyday AI invocations. The endgame of this revolution is the seamless integration of compute into every crevice of human life, like water and electricity—and we are standing squarely at the center of that turning point, right now.