Nvidia customers ordering AI servers for delivery next year are getting a number, not a reassurance. Bloomberg reported this week that AI server prices, including systems built on Vera Rubin and Grace Blackwell, will rise more than 15 percent in many configurations, with the exact increase varying by chip generation. The buyers are covering the industry’s rising memory bill. Every company sizing a 2027 compute budget now has to plan for double-digit inflation on top of whatever unit growth it already assumed.
Independent analyst newsletter StreetSignal used the news to test a question it had already been modeling: does Nvidia’s gross margin actually move when high-bandwidth memory (HBM) gets more expensive. Its answer, built from public filings and third-party cost estimates rather than anything Nvidia has disclosed, is that Nvidia has structured its pricing to protect gross-profit dollars, not gross-margin percentage, and that the distinction is about to matter more.
The mechanics start with Epoch AI’s teardown of the B200 accelerator, which StreetSignal cites as its baseline. Epoch estimates the B200 costs roughly $6,400 to build, with its 192GB of HBM3E memory priced around $15 per gigabyte. HBM makes up an estimated 45 percent of that bill of materials, StreetSignal calculates, yet only about 8.2 percent of what Nvidia charges for the finished chip. That gap is why a memory-cost spike that would sink a server maker’s margins barely dents Nvidia’s.
Run the arithmetic on a reported 20 percent HBM price increase, the figure StreetSignal attributes to Samsung and SK Hynix for 2026 deliveries, and the added cost per B200 comes to about $576, by the newsletter’s estimate. Passing that dollar-for-dollar to customers keeps Nvidia’s gross-profit dollars whole. It does not keep the percentage whole, because the added revenue carries no incremental margin of its own. StreetSignal calculates that holding an 81.7 percent chip-level margin against that cost increase would require roughly $3,150 in additional price, meaning every incremental memory dollar needs about $5.47 of price increase behind it to leave the margin ratio unchanged.
At the company level, the newsletter runs a sensitivity rather than a forecast: against Nvidia’s reported FY26 revenue of $215.9 billion and cost of revenue of $62.5 billion, if HBM represents 20 to 33 percent of total cost of goods sold, a fully absorbed 20 percent HBM increase would cost the company an estimated 145 basis points of gross margin at the midpoint, enough to pull a 75 percent run rate into the mid-73s. Nvidia does not break out HBM as a share of COGS, so these percentages are StreetSignal’s estimates, not company figures. Nvidia’s actual results have held near 75 percent through Q1 FY27, with CFO Colette Kress telling analysts on the Q3 FY26 call that manufacturing efficiencies, faster production cycles and a richer system mix would carry the load of keeping margins near 75 percent, and CEO Jensen Huang adding that Nvidia locks in memory supply far ahead of need. The company closed FY26 with $95.2 billion in manufacturing and supply commitments, up from $30.8 billion the year before.
The harder test sits in FY28. B200 carries 192GB of HBM3E; Rubin moves to 288GB of HBM4, roughly 50 percent more memory before any per-gigabyte premium, and Rubin Ultra could carry close to a terabyte, per StreetSignal’s estimate. More memory content as a share of each accelerator means Nvidia needs a bigger markup, not just a pass-through, to hold the same margin percentage, at the same moment AMD and custom ASICs are giving buyers a price reference they did not have before. Nvidia reports fiscal second-quarter results on August 26. Operators building 2027 and 2028 compute plans should treat the mid-70s margin commentary as a signal to watch, not a guarantee, and model server costs assuming memory content, not just memory price, keeps climbing.
Based on analysis published by StreetSignal (streetsignal.substack.com) on August 22, 2026.
