Jensen Huang has made bold predictions before. But the one he delivered on September 17, 2026, standing outside an AI summit convened by King Charles III at Dumfries House in Cumnock, Scotland, is now rippling through memory chip factories in South Korea, a new Nvidia-Palantir supply chain product, and a debate over whether “twice as many chips” means what Wall Street thinks it means.
“I expect Nvidia to sell twice as many chips next year as we do this year,” Huang told reporters, according to CNBC. He added: “The reason for that is because AI has so much contribution to the benefits of different industries, different economies, and you can see that in almost every single country that we’re in, people want to invest in AI.” Nvidia shares climbed as much as 2.8% in the hours after the remarks spread, a reaction later detailed by industry outlet MarketScale.
A week later, the story has split into two threads that matter more than the headline number. One is about who actually benefits if Nvidia doubles its chip output: mostly Samsung Electronics and SK Hynix, the two companies that make the high-bandwidth memory (HBM) every AI accelerator needs. The other is about whether Nvidia can even hit that number, given that its own chief financial officer has framed doubling as the best-case, supply-unconstrained scenario rather than a locked-in plan.
What Huang Actually Said, and Where
The setting matters. Huang wasn’t on an earnings call or a keynote stage. He was speaking to reporters ahead of an AI-focused gathering hosted by Britain’s King Charles III, an event that also drew representation tied to Google DeepMind, OpenAI, and Anthropic, according to reporting picked up by outlets including The Next Web. That’s an unusual venue for a chip-shipment forecast, and it’s part of why the remark traveled so fast: it read less like guidance and more like a spontaneous, confident aside.
Coverage landed within hours across CNBC, Bloomberg, South Korea’s Seoul Economic Daily, and the Chosun Ilbo. By September 22, follow-up pieces were already reframing the comment as a message aimed squarely at shareholders rather than a passing quote to a press pool. Nvidia has not issued a formal press release restating the forecast, which means every subsequent analysis, including this one, is working from the same handful of on-the-record sentences.
“Twice as Many Chips” Is Not the Same as Double the Revenue
The most important clarification, and the one that got lost in the first wave of headlines, is that Huang was talking about unit volume, not dollars. Nvidia’s chip business spans far more than data-center GPUs. It includes CPUs, networking and optical interconnect silicon, notebook processors, and Jetson robotics modules. A doubling in total unit count could be driven heavily by cheaper, high-volume parts rather than by Blackwell or Vera Rubin GPUs, the products actually gated by memory and power constraints.
That distinction matters because Nvidia has already put out a formal number that’s easier to hold it to: roughly 70% revenue growth guidance for the fiscal year ending January 2028, a figure CNBC has pegged against a market-implied revenue estimate of about $673 billion. Seventy percent growth is enormous by any normal company’s standard, but it isn’t a doubling. If Huang’s “twice as many chips” comment and the official 70% growth guidance both hold true at once, it implies the mix is shifting toward higher unit volumes at lower average prices, exactly what you’d expect if CPUs, networking gear, and notebook chips are pulling more weight in the count.
Huang’s CFO Already Told You the Catch
Nvidia CFO Colette Kress has previously described the doubling scenario as the supply-unconstrained case, according to the MarketScale analysis of the company’s recent statements. That’s corporate shorthand for: this is what happens if nothing breaks. And in a supply chain that depends on thousands of component suppliers delivering millions of parts on schedule, something usually breaks. A rack-scale AI system needs coordinated delivery of compute, memory, networking, cooling, and power hardware. A shortfall in any single category, most often HBM, can gate an entire rack shipment regardless of how many GPUs are sitting in a warehouse.
That framing turns Huang’s comment from a demand story into an operations story. Demand for AI compute hasn’t been in question for two years. What’s genuinely uncertain is whether the physical supply chain, memory fabs, substrate suppliers, cooling vendors, can be coordinated tightly enough to turn that demand into doubled shipments within twelve months.
From a $1 Trillion Backlog to a 2x Forecast: The Timeline
Huang’s Scotland comment didn’t come out of nowhere. At Nvidia’s GTC conference in March 2026, he told the audience that combined orders for Blackwell and Vera Rubin systems could total $1 trillion through 2027, an order-value figure rather than a unit or revenue number. Separately, Huang has said Nvidia shipped 6 million Blackwell GPUs over the prior four quarters, a shipment statistic that predates and is distinct from the new doubling claim. None of the currently available reporting ties that 6 million figure to a formal 2026 shipment target, so it can’t be used to say whether Nvidia beat or missed earlier guidance, only that the company has a track record of citing large, headline-friendly numbers at major public moments.
| Date | Forecast or Figure | What It Actually Measures | Source |
|---|---|---|---|
| March 2026 (GTC) | $1 trillion | Cumulative order value for Blackwell and Vera Rubin systems through 2027 | Nvidia GTC keynote |
| 2026 (reported) | 6 million units | Blackwell GPUs shipped over the prior four quarters | Huang public remarks |
| Most recent quarter | ~70% growth | Guided revenue growth for fiscal year ending January 2028 | Nvidia earnings guidance, via CNBC |
| Most recent quarter | ~$673 billion | Market-implied fiscal 2028 revenue estimate | CNBC |
| September 17, 2026 | 2x (doubling) | Total chip unit volume in 2027 versus 2026, across all product lines | Huang remarks, Scotland AI summit, via CNBC |
Laid out this way, the pattern is consistent: Nvidia’s public forecasts keep getting bigger and less precise about which metric they actually describe. Order value, shipment counts, revenue growth, and unit-volume doubling are four different measurements, and coverage this week has repeatedly conflated them. Readers comparing this story to Nvidia’s record $96.2 billion earnings quarter earlier this year should note that quarter measured revenue, not the unit counts Huang cited in Scotland.
The Real Constraint Is Memory, Not Compute
How Much HBM Nvidia Actually Needs
The supply-side reality behind Huang’s forecast comes down to one component: high-bandwidth memory. A Morgan Stanley analysis cited in September 2026 reporting estimates that Nvidia, Alphabet, and AMD combined will consume roughly 85% of global HBM supply in 2027, with Nvidia alone accounting for about 37.3% of that total. HBM has also grown from roughly 20% of a GPU’s material cost to more than 50% today, according to the same analysis. That’s the clearest evidence that Nvidia’s doubling ambition lives or dies on memory availability, not on its own fabs or TSMC’s wafer output.
Samsung, SK Hynix, and Micron’s 2027 HBM production capacity is already described as effectively sold out, with no meaningful new capacity expected until late 2027 or 2028 at the earliest. That’s a hard ceiling sitting directly under a forecast that assumes output can double within a similar window. Readers who followed the site’s earlier coverage of China’s AI chip prices jumping 50% on HBM shortage will recognize the pattern: memory scarcity, not GPU design or fab capacity, is the binding constraint across the entire AI hardware industry right now.
| Company | Estimated Share of 2027 Global HBM Supply | Role |
|---|---|---|
| Nvidia | ~37.3% | Largest single buyer, GPU maker |
| Alphabet | Part of combined ~85% | TPU and custom silicon buyer |
| AMD | Part of combined ~85% | Instinct GPU buyer |
| All other buyers | ~15% | Remaining global demand |
Note that the Alphabet and AMD figures are only available as part of the combined 85% estimate; Morgan Stanley’s analysis, as reported, breaks out Nvidia’s individual 37.3% share but does not further split the remaining allocation between the other two buyers.
Samsung and SK Hynix Are the Quiet Winners
If there’s a clean winner from Huang’s comment, it’s South Korea’s memory industry. Seoul Economic Daily reported that Samsung Electronics and SK Hynix shares rose in Korean trading immediately after the doubling forecast circulated, as investors read it as confirmation that HBM demand isn’t slowing down. That’s a notable reversal in framing: for most of 2026, memory shortages have been covered as a cost problem squeezing device makers and consumers. This is one of the first times the same shortage has been reported as a straightforward earnings tailwind for the two companies actually producing the memory.
It also reinforces a theme this site has tracked closely: memory, not logic, has become the scarce resource that determines who profits in the AI hardware cycle. Samsung previously detailed its 8-layer HBM4E memory built for Nvidia, running at 18Gbps per pin, and that product roadmap is now directly tied to whether Huang’s 2027 unit forecast is achievable at all. Every additional GPU Nvidia wants to ship needs a matching allocation of Samsung or SK Hynix memory stacks, and that allocation is reportedly locked in well past 2027.
Nvidia’s Answer: Put AI to Work on Its Own Supply Chain
A week before Huang’s Scotland remarks, on September 10, 2026, Nvidia and Palantir announced a sovereign AI collaboration explicitly aimed at supply chains, starting with Nvidia’s own operations. The stack combines Nvidia’s Nemotron open models with Palantir’s Foundry platform and its Artificial Intelligence Platform (AIP), grounded in Palantir’s Ontology layer, which maps an organization’s data, assets, processes, and decision rules. According to the companies, the first deployment is designed to spot supply constraints earlier and improve how materials get allocated across production, using Nvidia itself as the pilot customer before the product reaches outside buyers.
The two companies have said they intend to extend the approach to Palantir customers in manufacturing, retail, and technology. The architecture is designed to ship as a sovereign reference design that can run on-premises using Dell and Cisco hardware, or through colocation and cloud environments via Rackspace and Nebius. Read alongside the doubling forecast, the timing isn’t a coincidence: Nvidia becoming the first customer of a platform it plans to resell is a strong signal that internal supply chain visibility, not additional GPU design work, is what the company sees as the next constraint to remove.
Where AMD and Custom Silicon Fit Into the Picture
Nvidia isn’t the only company competing for the same scarce HBM pool. AMD, whose stock recently crossed the $1 trillion market cap mark for the first time, is folded into that same Morgan Stanley estimate of combined 85% HBM consumption alongside Nvidia and Alphabet. Custom silicon programs at hyperscalers add further pressure on the same memory suppliers, even though publicly available reporting doesn’t yet break out AMD’s Instinct-series shipment targets or Broadcom’s custom AI ASIC volumes in a way that allows a direct, apples-to-apples comparison with Nvidia’s doubling claim.
What is clear is that the competitive question in AI hardware has quietly shifted. It’s no longer primarily about whose GPU architecture benchmarks faster. It’s about which company can lock in enough HBM allocation from Samsung, SK Hynix, and Micron to actually build the racks its customers have already ordered. That’s a procurement and contracting fight as much as an engineering one, and it plays out mostly in supplier agreements that don’t get press releases.
What This Means for Enterprise AI Buyers
For companies planning AI infrastructure purchases into 2027, the practical takeaway isn’t about Nvidia’s headline number at all. It’s about allocation. MarketScale’s analysis for enterprise technology buyers highlights three consequences worth planning around: heavier competitive allocation pressure on GPU orders, rising networking and infrastructure costs that ride alongside GPU spending rather than being absorbed by it, and the reemergence of on-premises architectures as a first-class option rather than a fallback for companies that can’t get cloud capacity.
That last point is a meaningful shift in enterprise IT strategy. For much of the past two years, renting AI compute from a hyperscaler was framed as strictly simpler than buying and operating hardware. A tightening allocation environment, where even Nvidia’s own supply is described as constrained by its CFO, changes that calculus. Buyers who can secure guaranteed on-prem hardware allocations, even at a premium, may end up with more predictable AI roadmaps than those competing for cloud GPU capacity in a market where the biggest customer in the room is telling reporters it wants to double its own order book.
Historical Context: A Pattern of Big Numbers, Vague Denominators
Nvidia’s public messaging over 2026 has followed a consistent shape: large, attention-grabbing figures delivered in informal settings, later parsed by financial media into more precise but less exciting categories. The $1 trillion order figure at GTC in March was an order-value number, not a shipment or revenue commitment. The 6 million Blackwell units shipped figure was a historical shipment count, not a forward guide. Now the doubling forecast is a unit-volume projection spanning products far beyond the data-center GPUs most coverage pictures when it reads “chips.” Each of these numbers is real and sourced, but each measures something different, and readers who stack them together without checking the units end up with a distorted picture of Nvidia’s actual growth trajectory.
This isn’t unique to Nvidia. Big semiconductor and AI infrastructure companies have increasingly used investor-day and conference remarks to set expectations informally, ahead of the more conservative numbers that show up in SEC filings and formal earnings guidance months later. Huang’s Scotland comment fits that pattern closely: newsworthy, quotable, and directionally bullish, but not a substitute for the fiscal 2028 guidance Nvidia has already filed with regulators.
Predictions: What Happens Next
Based on the sourced reporting available as of September 25, 2026, a few outcomes look likely over the next two to four quarters:
- HBM pricing pressure will keep climbing through 2027, since Samsung, SK Hynix, and Micron capacity is already described as sold out with no meaningful new supply until late 2027 or 2028.
- Nvidia will likely announce additional supply chain visibility tools or partnerships, following the Palantir template, as it tries to close the gap between its unit-volume ambitions and what its component suppliers can physically deliver.
- Expect financial media and analysts to keep separating Huang’s unit-volume comments from Nvidia’s formal revenue guidance, since the two numbers already imply different growth stories once you account for product mix.
- Samsung and SK Hynix are positioned to keep benefiting from AI memory demand commentary in the near term, even as the broader memory shortage continues to raise costs for phone and laptop makers elsewhere in the market.
- Watch Nvidia’s next formal earnings call closely: if management repeats the doubling language with specific unit or revenue figures attached, it becomes real guidance. If it stays confined to conference remarks, treat it as directional confidence rather than a commitment.
The Bottom Line
Jensen Huang’s comment in Scotland was real, sourced, and consistent with a company that has spent two years riding AI demand to record results. But the more important story underneath it isn’t Nvidia’s confidence, it’s the fact that Nvidia’s own CFO has already flagged supply, not demand, as the deciding factor in whether that confidence pans out. Every additional chip in that doubling forecast needs a matching HBM allocation that Samsung, SK Hynix, and Micron say they can’t meaningfully expand until 2028. That makes this less a story about Nvidia’s ambitions and more a story about three memory companies in South Korea and the US quietly holding the pen on how fast the entire AI hardware industry is allowed to grow.
Frequently Asked Questions
Did Jensen Huang really say Nvidia will sell twice as many chips next year?
Yes. Speaking to reporters on September 17, 2026, ahead of an AI summit in Scotland, Huang said he expects Nvidia to sell twice as many chips in 2027 as it did in 2026, citing rising AI investment across industries and countries, as reported by CNBC.
Does “twice as many chips” mean Nvidia’s revenue will double?
No. Huang’s comment referred to unit volume across Nvidia’s full product range, including GPUs, CPUs, networking chips, notebook processors, and Jetson modules, not revenue. Nvidia’s own guidance points to roughly 70% revenue growth for the fiscal year ending January 2028.
What is actually limiting Nvidia’s ability to double chip output?
High-bandwidth memory supply. Nvidia’s own CFO has described the doubling scenario as the supply-unconstrained case, and Samsung, SK Hynix, and Micron’s 2027 HBM capacity is already reported as effectively sold out.
Who benefits most from Nvidia’s forecast?
Samsung Electronics and SK Hynix appear to be the clearest beneficiaries. Their shares rose in Korean trading after the forecast circulated, as investors interpreted it as confirmation of sustained HBM demand.
What is the Nvidia-Palantir supply chain deal, and is it related?
Announced September 10, 2026, the deal combines Nvidia’s Nemotron open models with Palantir’s Foundry and AI Platform to give companies earlier visibility into supply constraints. Nvidia is piloting it on its own operations before extending it to manufacturing, retail, and technology customers, a move widely read as Nvidia trying to remove its own supply chain bottlenecks ahead of its doubling target.
How does this compare to Nvidia’s $1 trillion GTC forecast from March 2026?
They measure different things. The GTC figure was cumulative order value for Blackwell and Vera Rubin systems through 2027. The September doubling comment is a unit-volume forecast for 2027 versus 2026 across Nvidia’s entire chip lineup. Both are real, sourced statements, but they aren’t directly comparable numbers.
Are AMD and other chipmakers facing the same memory constraint?
Yes. A Morgan Stanley analysis cited in September 2026 reporting estimates Nvidia, Alphabet, and AMD together will consume about 85% of global HBM supply in 2027, meaning all three face the same underlying memory allocation pressure, even though public shipment-volume comparisons between them aren’t currently available.
Should enterprise AI buyers change their purchasing plans because of this forecast?
Industry analysis suggests buyers should plan for tighter GPU allocation, higher networking and infrastructure costs alongside compute spending, and consider on-premises hardware as a serious option rather than defaulting entirely to cloud capacity, given that even Nvidia’s own supply is described as constrained.
