The most important disclosure from Nvidia’s record-setting earnings call on August 26, 2026, was not the $96.2 billion in quarterly revenue or the $108 billion guidance for Q3. It was actually a sentence from CFO Colette Kress that enterprise procurement teams and data center planners need to read before they finalize their 2027 infrastructure budgets: customer forecasts point to doubling next year, and Nvidia’s own guidance holds at 70% because the supply chain cannot keep up.
As Tech Times reports, that gap — between what buyers want and what memory manufacturers can deliver — is what makes AI server prices rise 15% or more on systems shipping in early 2027 and what will keep them elevated for at least 16 more months.
The constraint at issue is high-bandwidth memory — the ultra-fast, stacked chip architecture that feeds data to Nvidia’s AI accelerators — along with the advanced semiconductor packaging process required to assemble those chips alongside GPU dies. Together they form a dual-bottleneck structure that no single capital investment decision can quickly dissolve, and Nvidia’s own supply commitment figures confirm it: the company’s supply commitments more than doubled in a single quarter, from $119 billion at the end of Q1 to $279 billion at the end of Q2 of fiscal 2027, primarily driven by memory procurement for the Vera Rubin platform.
Record Quarter, Honest Warning
Nvidia’s second fiscal quarter of 2027, ended July 26, 2026, produced results that would look extraordinary from almost any company in any industry. Total revenue reached $96.2 billion, up 18% from the previous quarter and up 106% year-over-year — a performance reflected in record quarterly revenue and earnings that beat Wall Street consensus by approximately $4 billion. Data center revenue hit $89 billion, up 117% from a year prior. Non-GAAP earnings per diluted share of $2.22 beat the $2.09 analyst consensus. For the third fiscal quarter, the company provided a Q3 guidance of $108 billion in revenue — what will be its first quarter above the $100 billion threshold, guided with no China data center compute revenue assumed.
Kress acknowledged what many investors were already watching in the gross margin line. “Many of you have expressed concerns regarding our gross margins, as component costs have risen significantly,” she said. “We are experiencing extreme pricing conditions in memory. The magnitude of the price increase has memory pricing exceeded prior expectations and are headed even higher into next year.” The consequence is a guided decline across three consecutive quarters: from 75% in Q2, down to approximately 74% in Q3, to a gross margin trough of 71-72% in Q4 of fiscal 2027, before a recovery to 72% to 73% as Nvidia’s own price increases on its products take effect in fiscal 2028.
Kress did not frame this as a cost shock without a compensating benefit. She argued that tighter memory supply as a symptom of the same demand surge driving Nvidia’s own growth differs fundamentally from a pure cost hit: “Unlike a component that simply raises our cost with no offset benefit, tighter memory supply is a symptom of the same demand surge that is driving our own growth,” she said. CEO Jensen Huang was more direct: unconstrained demand would produce growth “a lot higher” than the 70% Nvidia is guiding. Nvidia shares rose approximately 6% in after-hours trading following the report.
How Does HBM Create a Supply Ceiling?
To understand why Nvidia cannot simply produce more output in response to higher demand, it helps to understand what makes high-bandwidth memory different from every other component in an AI server — and why that difference creates a ceiling that is structural, not temporary.
Standard server DRAM delivers data at roughly 50 to 100 gigabytes per second per memory module. HBM3E, the type used in Nvidia’s current Blackwell generation, HBM3E delivers 1.2 terabytes per second of bandwidth per stack by stacking multiple DRAM dies vertically and connecting them through microscopic copper pathways called through-silicon vias (TSVs), then bonding that stack directly alongside the GPU on a silicon interposer. The result is dramatically higher bandwidth — but at a manufacturing cost that compounds the supply constraint in two independent ways.
The first bottleneck is wafer area. Producing one gigabyte of HBM consumes roughly three to four times the wafer area per gigabyte of HBM as standard DRAM requires. Every fab wafer reallocated from commodity DDR5 production to HBM production therefore removes three to four times as many conventional memory bits from the supply pool as it adds in HBM bits. HBM already accounts for roughly HBM drives 30% DRAM revenue despite representing only about 8% of total memory output in bits — and that ratio is still widening as Vera Rubin’s demand ramps.
The second bottleneck is TSMC’s CoWoS (Chip-on-Wafer-on-Substrate) advanced packaging capacity. Producing HBM is not enough — each stack must be bonded to the GPU die on a silicon interposer using CoWoS, a specialized packaging technology that TSMC operates at a limited number of dedicated facilities. TSMC’s CoWoS lines are estimated to be running at TSMC CoWoS capacity at maximum utilization in 2026, with Nvidia consuming roughly 60% of available capacity for its data center accelerators. Even if Samsung and SK Hynix were to produce significantly more HBM4 dies tomorrow, TSMC’s CoWoS lines could not package and deliver them into AI servers at a faster rate. These two constraints — HBM die production and CoWoS packaging — compound each other, which is why supply remains structurally limited regardless of how aggressively the memory makers invest.
The Vera Rubin architecture intensifies this pressure further. Each Vera Rubin GPU carries 288 gigabytes of HBM4 — a 50% increase over the 192 gigabytes of HBM3E on the Blackwell B200. Vera Rubin Ultra, the follow-on platform, targets the Vera Rubin Ultra HBM4E target spec of up to one terabyte of HBM4E per GPU. With Vera Rubin already in full production and ramping across CoreWeave, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure, the demand signal into the memory supply chain is not merely strong — it is accelerating into each successive GPU generation that carries more memory per chip than the last.
Nvidia’s balance sheet shows an inventory balance of $32 billion at Q2’s close, as the company pre-positioned Vera Rubin components at scale — a figure that itself reflects the memory procurement required for the ramp. That pre-positioning itself tightens near-term availability for other buyers in the broader memory ecosystem.
What Enterprise Buyers Are Actually Paying
The memory squeeze is not simply an abstract constraint on Nvidia’s growth rate. It is a concrete cost event for every organization buying AI infrastructure for 2027.
Bloomberg reported, and multiple outlets including Tom’s Hardware and The Next Web confirmed, that some of Nvidia’s largest customers and the server companies that build systems for major data center operators — including Microsoft, Google, and Oracle — have been notified that server prices rise more than 15% for Vera Rubin and Grace Blackwell systems shipping in early 2027 in many configurations. The exact magnitude of each increase depends on the chip generation and memory configuration, with systems carrying maximum memory — typically the configurations used for the largest AI models — facing the sharpest increases.
For organizations planning large-scale AI infrastructure procurement, the practical implication is straightforward: systems that fit 2026 budget models do not have 2027 delivery prices. A 15% increase on AI server systems that already run into the hundreds of thousands of dollars each translates into material additional CAPEX for any facility sized around current hardware economics. Enterprise buyers whose deployment windows extend into early 2027 should treat that price increase as a planning input, not a contingency.
Nvidia has been explicit that it intends to protect gross profit dollars by passing memory cost increases through to customers rather than absorbing them at the margin line. The company’s estimated revenue opportunity per gigawatt of capacity has expanded from roughly $18 billion during the Hopper generation to $25 billion with Blackwell, and is estimated at approximately $40 billion with Vera Rubin when accounting for GPUs, CPUs, networking, systems, and software. Each generation that expands the per-rack sale bundle gives Nvidia more pricing flexibility even as individual component costs rise.
How Might This Affect AI Development Timelines?
Huang used the earnings call to frame the demand environment in terms that go beyond a single company’s outlook. In Huang’s AI inflection point quote from the earnings release, he said: “AI has reached its inflection point. This time last year, one lab alone was driving the buildout. Today, we have a golden age of new AI labs and startups, multiple frontier labs scaling in parallel, a thriving open-model ecosystem and physical AI coming online — with strong momentum across the U.S. and around the world.”
For memory makers, Nvidia’s candor is exceptional forward visibility in an industry not known for it. For customers, it is a call to act on procurement decisions before early 2027 pricing takes full effect — or to plan for a higher cost basis that, on current evidence, will not resolve until at least the end of fiscal 2028.
—
Photo Credit: DANIEL CONSTANTE / Shutterstock.com
