Bloomberg: AI Server Prices to Rise More Than 15% on Early-2027 Shipments as Memory Costs Climb
Bloomberg reported August 22 that buyers of servers built around Nvidia's Vera Rubin and Grace Blackwell chips were told to expect increases above 15% on systems shipping early next year, four days before Nvidia's quarterly earnings report.
The Notice Nobody Sent
On August 22, 2026, Bloomberg reported that buyers of the biggest AI servers had gotten word: prices are going up, in many cases by more than 15%[1]. The increase hits systems shipping in early 2027, and it covers machines built around Nvidia's newest chips — the Vera Rubin generation and the Grace Blackwell generation[1]. How much any one buyer pays depends on which chip they order and how much memory goes into the box[1].
Almost every headline that followed read some version of "Nvidia customers notified." That phrasing is accurate, but it hides a detail buried in Bloomberg's own reporting. The notices didn't come from Nvidia. They came from the contract manufacturers who assemble servers for Microsoft, Google and Oracle, telling their own customers what the new numbers will be[1][3]. Nvidia was asked for comment and didn't respond[3]. It has not confirmed any change to its own list prices.
That distinction matters more than it looks like it should, because it changes who the story is actually about.
Where the Money Is Actually Going
The stated reason for the price jump isn't Nvidia's processors. It's the memory chips packed around them. DRAM is the fast, short-term memory a server uses to hold data the processor is actively working on, and its price has gone almost straight up. TrendForce reported in January that Samsung and SK Hynix were pushing server DRAM contract prices up 60% to 70%[6]. Korean business press put the ask as high as 70%[8]. J.P. Morgan's research arm estimates DRAM prices will have risen more than 400% from the start of 2024 to the end of 2026[10].
In some AI server configurations, memory now makes up 40% to 50% of what it costs to build the machine[16]. That's why a modest-sounding price increase on the finished server can still represent a real financial squeeze passed down the chain. Every side in this story agrees memory got expensive. What they don't agree on is who should eat the cost, and what the timing means.
Nvidia reports its quarterly earnings on Wednesday, August 26 — four days after this report surfaced[16]. The company has guided investors to expect around $91 billion in revenue[16]. Analysts are watching whether it can hold gross margins in the mid-70% range as it shifts production toward the newer Rubin chips. Bank of America estimates the memory cost headwind alone could shave about 0.6 of a percentage point off that margin[16]. Whoever ends up absorbing the higher memory cost — Nvidia, the contract builders, or the cloud companies — is the one who takes that hit.
The Case for Nvidia
Nvidia's strongest argument is simple arithmetic. A modern AI server isn't just a processor. It's a processor surrounded by a large volume of memory bought on the open market, at whatever the market is charging that quarter. If memory already accounts for 40% to 50% of what a system costs to build[16], and memory prices have roughly doubled in some categories[14], then a 15% increase in the finished system price is actually less than a full pass-through. Under this reading, somebody in the chain — Nvidia or its builders — is absorbing part of the hit, not just forwarding all of it.
There's also a supply argument. Nvidia competes with every other electronics maker for the same limited pool of memory wafers. If it doesn't pay up, it doesn't get the memory, and it ships fewer systems. Paying more, in this framing, is what keeps the pipeline moving at all.
The Case Against Absorbing It Twice
Buyers like Microsoft, Google and Oracle — and the contract manufacturers standing between them and Nvidia — see it differently. They're already paying for the memory shortage once, in their own component purchases. Meta has raised its 2026 capital spending floor to $130 billion to $145 billion, citing memory price inflation as a factor[9]. Microsoft has attributed $25 billion of its own spending increase to higher component costs[9]. From their chair, getting hit again on the finished server price means paying for the same shortage twice.
The contract server builders occupy the thinnest position in this fight. They run on the smallest margins of anyone in the chain, with the least room to absorb a surprise cost increase themselves[1]. That's also likely why they're the ones actually sending these notices — and why they're almost invisible in the coverage. Cloud providers, meanwhile, sell AI computing power on multi-year contracts that assume the cost of hardware keeps falling over time, the way it usually does in tech. A sudden 15% jump on servers breaks that assumption in a way a chart alone won't capture. Dell'Oro Group now projects 2026 data center capital spending will top $1 trillion, with the top four U.S. cloud providers increasing spending 78%[9].
Seoul Sees a Different Story
Flip the story to South Korea, and the same price increase looks like something else entirely: overdue payback. Samsung and SK Hynix spent years selling memory chips below cost after previous cycles of overbuilding capacity[8]. Korean business outlet KED Global described the current moment as a "decisive shift to a seller's market[8]." In that framing, high prices aren't extraction — they're the signal that finally justifies building the next generation of fabrication plants, which cost enormous sums and take years to complete.
There's a physical constraint underneath all of this that nobody disputes. High-bandwidth memory, the specialized kind used in AI chips, eats up roughly three times the wafer capacity per unit that standard memory does. That share of total DRAM wafer production is projected to rise to 23% in 2026, up from about 19%[13]. Every wafer devoted to an AI server is a wafer that isn't going into a laptop or a phone. New factories take years to build, so in the short term, no pricing decision — generous or stingy — changes how much memory physically exists. SK Hynix alone is reportedly spending at a scale of $720 billion to try to close that gap[18].
That scarcity has a spillover cost. PC and phone makers, who never bid for AI memory in the first place, are competing for what's left of standard DRAM supply, and prices there have followed AI memory upward[13]. Deloitte has argued the broader memory crunch may not ease until 2029[11].
What the Headlines Left Out
Coverage of the same report varied by what each outlet chose to foreground. Bloomberg's own headline led with "Nvidia Customers Notified," a passive construction that lets readers assume Nvidia sent the notice, even though its own story attributes the notices to contract server builders[1]. Reuters flagged the secondhand sourcing explicitly in its headline and noted Nvidia's non-response, which is about as disciplined as wire coverage gets[3]. CNBC's version used the word "warned" instead of "notified," a slightly heavier verb pitched toward an investor audience ahead of the earnings report[2]. Fortune's headline leaned on the Vera Rubin and Grace Blackwell product names, which reads more like a Nvidia product story than a supply-chain cost story[4]. South China Morning Post ran it comparatively straight, without its usual China or export-control framing[5]. And The Register's January piece on the same underlying memory shortage used the phrase "Mem-ageddon" and described an AI "frenzy" set to "wallop" DRAM prices — loaded language layered on top of accurate wafer-allocation numbers[13].
Nvidia reports earnings August 26. Whatever the company says then — or doesn't say — about memory costs and margins will be the next data point in a story that, so far, has been told almost entirely by everyone except Nvidia itself.
Summary
On Saturday, August 22, 2026, Bloomberg News reported that some of the biggest buyers of AI servers had been told prices will rise more than 15% in many cases[1]. The increases apply to systems shipping early next year — that is, early 2027. They cover machines built around Nvidia's newest chips, the Vera Rubin generation and the Grace Blackwell generation[1]. How big the increase is depends on which chip generation a buyer takes and how much memory is packed into the box[1]. Nvidia did not respond to a request for comment[3].
One detail in the original report is easy to lose. Bloomberg says the notices came from the companies that build servers under contract for large data center operators such as Microsoft, Alphabet's Google and Oracle[1][3]. Those contract builders told their customers. Bloomberg's story does not describe Nvidia sending a price notice itself, and Nvidia has not confirmed a list-price change. Most follow-on headlines used the passive phrase "Nvidia customers notified," which leaves the sender unnamed.
The stated cause is memory. Not the AI processor itself, but the memory chips bolted to it. DRAM is the fast working memory a server uses to hold data the processor is actively crunching. Its price has gone nearly vertical. TrendForce reported in January that Samsung and SK Hynix were pushing server DRAM contract prices up 60–70%[6], and Korean business press reported the ask ran as high as 70%[8]. J.P. Morgan Global Research estimates DRAM prices will have risen more than 400% from the start of 2024 to the end of 2026[10].
The genuine dispute is not whether memory got expensive — every side agrees it did. It is who should absorb it and what it signals. One camp reads a pass-through of real input costs, the ordinary working of a shortage. Another reads a company with enough market power to hand its cost problem to customers days before reporting earnings. A third, in Seoul and among memory investors, reads the whole thing as overdue payback for years of underinvestment in memory factories. Nvidia reports quarterly results on Wednesday, August 26[16].
The Event
On August 22, 2026, Bloomberg News reported that buyers of AI servers had been notified of price increases exceeding 15% in many cases, effective on systems shipped in early 2027[1]. The increases apply to systems built around Nvidia's Vera Rubin and Grace Blackwell chips, and vary by chip generation and memory configuration[1]. Bloomberg reported that companies building servers under contract for large data center operators including Microsoft, Alphabet's Google and Oracle had recently informed their customers of the coming increases[1][3]. Nvidia did not respond to a request for comment made outside regular business hours[3]. Reuters, CNBC, Fortune and South China Morning Post carried the report the same day[2][3][4][5].
Undisputed Facts
- Bloomberg News reported the price increases on August 22, 2026; the figure cited is "more than 15% in many cases," not a single fixed percentage[1].
- The increases apply to systems shipped starting early next year, and vary by Nvidia chip generation and by how much memory is in the configuration[1].
- Bloomberg attributed the notifications to contract server builders informing their customers, not to a direct Nvidia announcement; Nvidia did not comment[1][3].
- Vera Rubin and Grace Blackwell are Nvidia's current flagship AI system platforms and are named as affected[1][5].
- Memory prices rose sharply through 2026: TrendForce reported Samsung and SK Hynix seeking server DRAM increases of 60–70% in early 2026[6], and revised its conventional DRAM contract price increase estimate up to 90–95% in a June 3 update[14].
- Samsung and SK Hynix raised high-bandwidth memory (HBM3E) supply prices by roughly 20% for 2026[14].
- Dell'Oro Group raised its 2026 data center capital spending outlook above $1 trillion, citing AI buildouts and memory cost inflation[9].
- Nvidia is scheduled to report fiscal second-quarter 2027 results on Wednesday, August 26, 2026, with company guidance of about $91 billion in revenue, plus or minus 2%[16].
- Memory can account for 40–50% of total build cost in some AI server configurations[16].
The Pressure
Strip away the moralizing and blame. What structural realities persist regardless of which narrative wins?
- Wafers are the hard limit
- There is a fixed amount of memory-fab capacity in the world, and high-bandwidth memory for AI eats about three times the wafer area per unit that standard DRAM does. HBM's share of total DRAM wafer output rises to 23% in 2026 from about 19%[13]. New fabs take years. So in the short run, every AI server built is memory not available to a laptop, and no pricing decision changes that physical fact[11].
- Margin defense before an earnings print
- Nvidia reports on August 26 with roughly $91 billion in guided revenue and analyst attention fixed on whether gross margin holds in the mid-70s[16]. Whoever absorbs the memory cost takes the margin hit. That is the real object of the negotiation, regardless of how it is described publicly.
- Cloud contracts assume falling unit costs
- Hyperscalers sell AI compute on multi-year terms priced against an assumption that hardware gets cheaper per unit of work over time. A durable step up in system cost breaks that assumption, which is why buyers push back hard on a 15% increase that a chart would call modest.
- Memory's cycle memory
- Samsung, SK Hynix and Micron spent prior cycles selling below cost after overbuilding. That history makes them slow to add capacity and quick to hold price when leverage arrives. It is a structural reflex, not opportunism in the moment.
Material realityRegardless of who is blamed, three things are measurably true. Memory is scarce and expensive: DRAM prices are estimated to have risen more than 400% from the start of 2024 to the end of 2026[10], with server DRAM around $10 per gigabyte in 2026 and projected to ease toward roughly $5 by 2030[9]. AI infrastructure spending is enormous and still growing: Dell'Oro puts 2026 data center capex above $1 trillion, with the top four U.S. cloud providers up 78%[9]. And the shortage has a long tail: Deloitte argues it may not ease until 2029[11], while SK Hynix is spending at a scale reported at $720 billion to close the gap[18]. A 15% increase on early-2027 systems is a small number sitting on top of these very large ones. It does not, by itself, tell you whether AI buildout slows.
Narrative as a weaponThree actors are shaping how this lands. Nvidia, by saying nothing before its August 26 earnings, lets the memory makers be the named cause while its own list prices stay unconfirmed. The contract server builders — the parties Bloomberg actually identifies as sending the notices — are invisible in almost every headline, which suits them; they have the least pricing power and the most to lose from being named. Korean memory makers and their national press are actively framing the same price move as the correct functioning of a market that punished them for a decade. Meanwhile investors and short-sellers on both sides of the AI trade want this read as a signal: bulls as proof of demand so strong that suppliers can charge more, bears as the first crack in an unsustainable buildout. The single most useful correction for a reader is the smallest one: the reporting does not say Nvidia announced a price increase. It says buyers were told to expect one.
How Each Side Sees It
Each major actor’s view — how it frames things, its underlying incentive, and how it’s materially affected. Tap a side to read it.
Frames it asNvidia's strongest case is that this is arithmetic, not leverage. A modern AI server is not one chip. It is a processor surrounded by a very large amount of memory, and memory is bought on the open market at whatever the market charges. If memory is 40–50% of what a box costs to build[16], and memory contract prices roughly double[14], then a 15% rise in the finished system price is actually less than full pass-through — the builder is absorbing part of it. Nvidia would also argue that holding prices flat would mean building fewer systems, since it competes for the same scarce memory wafers as everyone else. Paying up for supply is how you keep shipping.
WhyProtect gross margin through a product transition. Analysts are watching whether Nvidia can hold margins in the mid-70s as the mix shifts to Rubin; BofA has put the memory headwind at about 60 basis points, or 0.6 of a percentage point off gross margin[16]. Passing the cost forward defends that number.
Impact on themNvidia reports earnings August 26 with guidance of about $91 billion in revenue[16]. CEO Jensen Huang has publicly described a potential $1 trillion in cumulative Blackwell and Rubin revenue from 2025 through calendar 2027[16]. Higher system prices support that revenue figure but raise the question of whether customers keep buying at the same unit volume.
Frames it asThe buyers' strongest argument is that they are already paying for this twice. They bear the memory inflation directly in their own component purchasing, and now again in the finished system price. Meta raised its capital spending floor to $130–145 billion citing memory price inflation, and Microsoft attributed $25 billion of its spending to higher component pricing[9]. Their case is that a cost increase should be shared across a supply chain, not routed entirely downstream to whoever signs the last check. The contract builders — the firms that actually assemble the machines — have the thinnest margins of anyone in the chain and the least room to absorb anything.
WhyKeep the cost per unit of AI compute falling, because their business model assumes it will. Cloud providers sell compute on multi-year contracts priced against an expectation that hardware gets cheaper per unit of work. A 15% step up in hardware cost cuts against that.
Impact on themThe top four U.S. cloud providers increased data center capital spending 78%[9]. Dell'Oro now puts 2026 data center capex above $1 trillion[9]. Higher server prices either raise what customers pay for cloud AI, or compress cloud margins, or slow the buildout — and which of the three happens is not yet visible.
Frames it asThe memory industry's strongest case is that it spent a decade being punished for building capacity. Memory is a boom-bust commodity; the last several cycles ended with these firms selling below cost. They argue that today's prices are the market finally paying enough to justify new fabrication plants, which take years and enormous capital. Building high-bandwidth memory for AI consumes about three times the wafer capacity of standard DRAM per unit shipped, so serving AI demand physically shrinks what is left for everything else. In their telling, high prices are the signal that summons new supply — and suppressing them would extend the shortage, not fix it.
WhyConvert scarcity into the profits and forward commitments needed to fund enormous fab construction. SK Hynix is in the middle of a buildout reported at $720 billion in scale[18]. Korean press described the moment as a "decisive shift to a seller's market"[8].
Impact on themHistoric pricing power. Samsung and SK Hynix sought server DRAM increases of up to 70%[8]; Reuters reported Samsung hiking memory prices by up to 60% as the shortage worsened[12]. HBM will consume 23% of total DRAM wafer output in 2026, up from about 19%[13].
Frames it asThis group's argument is that they are collateral damage in an auction they never entered. Memory makers have shifted wafer capacity toward high-margin AI memory, and what is left for ordinary DDR4 and DDR5 has become scarce and expensive[13]. They did not bid up the price and cannot pass it on as easily as a hyperscaler can. Apple and Qualcomm have already raised prices citing chip shortages. Their crux is different from everyone else's: for them the story is not Nvidia's margin, it is whether a laptop costs more next year.
WhySecure memory allocation at survivable prices and avoid losing consumers to sticker shock.
Impact on themServer DRAM average selling prices are projected around $10 per gigabyte in 2026, before easing toward roughly $5 per gigabyte by 2030[9]. Deloitte has argued the memory crunch is deeper than expected and may not ease until 2029[11].
Like this article?
The Bias Ledger average rating 3
The same story, as framed by outlets across the spectrum, ordered least to most biased. The bias score (1 = straight, 10 = heavily spun) is an AI assessment of that framing — click an outlet to see its track record. The tell is the word choice or omission that reveals the angle.
| Outlet | Vantage | Bias | How they frame it | The tell |
|---|---|---|---|---|
| Reuters | U.K.-based wire, center | 1 | "Nvidia customers notified about AI-related price hikes above 15%, Bloomberg News reports" — flags the second-hand sourcing in the headline itself and notes Nvidia did not respond to a comment request. | Minimal framing; the explicit "Bloomberg News reports" tag and the no-comment line are the disciplined choices. Adds no independent reporting of its own. |
| Fortune | U.S. center-left, business | 2 | Straight restatement of the Bloomberg report, naming Vera Rubin and Grace Blackwell in the URL and copy. | Product-name emphasis makes it read as a Nvidia product story. Little space given to memory makers as the origin of the cost, which leaves Nvidia as the apparent sole cause. |
| South China Morning Post | Hong Kong, owned by Alibaba Group | 2 | "Nvidia customers notified of AI-related price rises above 15%" — carried as a global tech-industry cost item. | Notably restrained; no China-angle overlay and no export-control framing, which is a departure from its usual Nvidia coverage. The omission is the tell — a cost squeeze on U.S. hyperscalers is left to speak for itself. |
| Bloomberg | U.S. center, financial | 3 | "Nvidia Customers Notified About AI-Related Price Hikes Above 15%" — sourced to people familiar, with the cause given as soaring memory costs. | The passive "Notified" hides who sent the notice. The body says contract server builders told their customers; the headline lets readers infer Nvidia did. Placing "Nvidia" first in a headline about a supply-chain pass-through puts the brand where the actor should be. |
| CNBC | U.S. center, business | 3 | "Nvidia customers reportedly warned about AI-related price hikes" — hedged with "reportedly," pitched to investors ahead of the August 26 earnings report. | "Warned" is a slightly heavier verb than "notified" and carries a hint of threat. The earnings-preview framing treats the story mainly as a signal about Nvidia's stock rather than about what data centers will cost. |
| KED Global | South Korean business press (Korea Economic Daily) | 4 | "Samsung, SK Hynix seek up to 70% server DRAM price hikes as AI boom tightens supply" — the same underlying event told from the supplier's chair, described as a shift to a seller's market. | The subject changes entirely. Here the memory makers are the protagonists winning a long-deserved cycle, not the cause of anyone's cost problem. "Seek" softens the ask; there is no downstream-pain framing at all. |
| The Register | U.K. tech trade press, skeptical house style | 6 | "Mem-ageddon: AI chip frenzy to wallop DRAM prices with 70% hike" — casts the AI buildout as inflicting damage on the rest of the computing market. | "Mem-ageddon," "frenzy" and "wallop" are three loaded words in one headline. The framing is accurate on the wafer-allocation numbers but editorializes the motive as mania rather than demand. |
References
- Nvidia Customers Notified About AI-Related Price Hikes Above 15% — Bloomberg · U.S. financial newswire owned by Michael Bloomberg; center, market-focused
- Nvidia customers reportedly warned about AI-related price hikes — CNBC · U.S. business network owned by Comcast/NBCUniversal; center, investor-oriented
- Nvidia customers notified about AI-related price hikes above 15%, Bloomberg News reports — Reuters · U.K.-headquartered wire service owned by Thomson Reuters; center
- Nvidia customers notified about AI-related price hikes above 15% — Fortune · U.S. business magazine; center-left editorially, business-friendly reporting
- Nvidia customers notified of AI-related price rises above 15% — South China Morning Post · Hong Kong daily owned by Alibaba Group; operates under Hong Kong's press environment
- Samsung, SK Reportedly Hike Server DRAM Prices 60-70% – Google, Microsoft in the Queue — TrendForce · Taiwan-based commercial market research firm; revenue comes from selling data to the semiconductor industry it covers
- Samsung, SK Hynix seek up to 70% server DRAM price hikes as AI boom tightens supply — KED Global · English edition of the Korea Economic Daily; South Korean business press, generally favorable to national industrial champions
- AI Infrastructure Buildouts and Memory Cost Inflation Drove Data Center Capex Higher in 1Q 2026 — Dell'Oro Group · U.S. commercial telecom/data-center market research firm; paid by vendors and operators for subscription research
- The AI-Driven Memory Shortage: DRAM Prices, Inflation and Market Risks — J.P. Morgan Global Research · Research arm of a major U.S. investment bank with trading and banking relationships across the semiconductor sector
- Why the memory chip crunch is greater than expected, and may not ease until 2029 — Deloitte · Global professional services firm; consulting clients include semiconductor and cloud companies
- Exclusive: Samsung hikes memory chip prices by up to 60% as shortage worsens, sources say — Reuters · U.K.-headquartered wire service owned by Thomson Reuters; center
- Mem-ageddon: AI chip frenzy to wallop DRAM prices with 70% hike — The Register · U.K. technology trade publication with an openly skeptical, irreverent house style
- Samsung, SK hynix Reportedly Plan ~20% HBM3E Price Hike for 2026 as NVIDIA H200, ASIC Demand Rises — TrendForce · Taiwan-based commercial market research firm selling data to the semiconductor industry it covers
- Nvidia Q2 Earnings Preview: Gross Margin and Rubin Hold Key to Breakout — TradingKey · Commercial investment-research and trading-content site; audience is retail traders
- An inside look at SK Hynix's $720 billion AI-fueled buildout that's taking over South Korea — CNBC · U.S. business network owned by Comcast/NBCUniversal; center, investor-oriented