Qualcomm-AWS $60B AI Chip Deal: Stock Jumps 7% [2026]

Qualcomm just became the second major chipmaker in six months to walk away from a multibillion-dollar hyperscaler deal that has nothing to do with smartphones. On September 8, 2026, Qualcomm confirmed a multi-year, multi-generational partnership with Amazon Web Services to co-develop custom AI inference silicon and high-speed optical networking gear for AWS data centers, according to Tech Times. Qualcomm stock jumped as much as 7% the same day, and the terms attached to the deal, a warrant for up to 25 million Qualcomm shares tied to as much as $60 billion in future purchases, mark the company’s biggest public step yet into the data center business it walked away from in 2018.

The Qualcomm AWS deal lands nine days before this article and less than three months after Qualcomm’s June 24, 2026 unveiling of its Dragonfly C1000 server CPU, when the company said it was building custom silicon for two more hyperscalers it hadn’t named yet. AWS is now confirmed as one of them, and it’s arguably the more important customer: the world’s largest cloud provider, and a company that has spent years building its own in-house Trainium and Graviton chips specifically to reduce its dependence on Nvidia.

What Qualcomm and Amazon Actually Signed

According to Qualcomm’s own 8-K regulatory filing, cited by Futurum Group, the agreement covers two distinct product lines: custom AI inference chips designed specifically for AWS workloads, and optical connectivity hardware capable of speeds up to 1.6 terabits per second. That optics component matters as much as the chip itself. Data centers running large AI clusters increasingly bottleneck on the networking fabric between accelerators, not just the accelerators themselves, and Qualcomm is positioning itself as a supplier for both halves of that problem.

The deal is structured as a warrant rather than a straightforward purchase order. Amazon received the right to acquire up to 25 million shares of Qualcomm stock at an exercise price of $161.26 per share, an arrangement TechTarget values at roughly $4 billion if fully exercised. An initial tranche of 3.75 million shares vests immediately based on binding commitments already in place. The remaining 21.25 million shares vest only if Amazon’s purchases of Qualcomm chips, optics, and manufacturing services reach a cumulative $60 billion before the warrant expires on September 3, 2036.

That distinction, a ceiling that unlocks equity rather than a signed purchase order, is the detail most outlets have been careful to flag. The Register notes that Amazon’s purchases and other commercial commitments under the agreement could total up to $60 billion in payments through September 2036, framing it explicitly as a ceiling rather than a guaranteed revenue stream for Qualcomm.

The $60 Billion Warrant Structure, Broken Down

Warrant-linked chip deals aren’t new in 2026’s AI infrastructure buildout, but the specific mechanics of the Qualcomm AWS deal are worth laying out on their own, since the headline number ($60 billion) and the actual near-term value (roughly $4 billion in stock) tell very different stories about how committed this partnership really is today versus over the next decade.

Deal Term	Detail	Source
Announcement date	September 8, 2026	Tech Times, The Register
Total warrant size	Up to 25 million Qualcomm shares	TechTarget, Futurum Group
Exercise price	$161.26 per share	TechTarget, Futurum Group
Approximate warrant value if fully vested	~$4 billion	TechTarget
Initial vesting tranche	3.75 million shares (immediate, tied to binding commitments)	Futurum Group
Full vesting condition	Cumulative purchases reaching $60 billion	The Register, Futurum Group
Warrant expiration	September 3, 2036	Futurum Group, The Register
Product scope	Custom AI inference silicon + up to 1.6 Tb/s optical connectivity	Futurum Group
Qualcomm stock reaction	Gained as much as 7% intraday on announcement day	Tech Times

The ten-year horizon on the warrant is itself a signal. Qualcomm isn’t promising AWS chips will ship next quarter, and AWS isn’t promising to buy $60 billion worth of anything on a fixed schedule. Instead, both companies structured an incentive: the more AWS actually buys, the more equity upside Qualcomm hands over, which aligns Qualcomm’s stock performance with the success of a product line that doesn’t exist yet in shipping form.

Why AWS Needed a Qualcomm Chip Deal

AWS is not short on AI chip options. It already designs its own Trainium training chips and Inferentia inference chips in-house, with Marvell serving as lead design partner on that silicon under a five-year agreement the two companies expanded in December 2024. AWS also remains, by a wide margin, Nvidia’s largest cloud customer, running enormous fleets of H100, H200, and Blackwell-generation GPUs for customers who specifically want Nvidia’s CUDA software stack.

So why add a third chip supplier? The practical answer is inference economics. Training large models still leans heavily on Nvidia GPUs and in-house Trainium silicon, but inference, the process of actually running a trained model to answer a query, is a different workload with different power and cost profiles, and it’s the workload growing fastest as AI products scale to hundreds of millions of users. Qualcomm’s pitch, built on its decades of low-power Snapdragon smartphone chip engineering, is that it can deliver inference silicon with a better performance-per-watt profile than general-purpose GPUs, a specialization Qualcomm already applied to its June 2026 Dragonfly C1000 server CPU and AI300 inference accelerator launch for Meta and Microsoft.

Nvidia’s dominance of the broader AI accelerator market remains enormous. Estimates for 2026 from named analyst firms cluster in a wide band: Silicon Analysts projects Nvidia’s revenue share falling to around 75% in 2026 after peaking near 87% in 2024, while IDC-linked estimates put the figure closer to 81%. Either way, Nvidia still controls somewhere between three-quarters and four-fifths of data center AI accelerator revenue, which is exactly the concentration risk hyperscalers like Amazon have been trying to hedge against with in-house silicon and outside custom-chip partners since 2024.

Qualcomm’s Long Road Back to the Data Center

Qualcomm spent eight years outside the data center business after shutting down its Centriq server CPU project in 2018. The Dragonfly C1000 launch on June 24, 2026, at Qualcomm’s Investor Day in New York, marked the company’s return, pairing a server CPU with a new AI300 inference accelerator. Meta signed on as the headline Dragonfly customer under what Qualcomm called a “multi-generation collaboration,” and Microsoft joined as a second named customer. At the time, Qualcomm said it was building custom silicon for two additional hyperscalers it declined to name.

AWS is now confirmed as one of those unnamed partners, roughly eleven weeks after the Dragonfly C1000 launch. That timeline matters for reading the AWS deal correctly: it isn’t a standalone announcement so much as the second major data point in a pattern Qualcomm set out publicly in June. If the company follows through on naming its fourth hyperscaler customer, that would put Qualcomm in data center silicon conversations alongside Meta, Microsoft, and AWS inside a single calendar year, a dramatic reversal for a company that had zero server chip revenue as recently as May 2026.

The Custom Silicon Field Qualcomm Is Entering

Qualcomm is a late arrival to a custom AI chip market that Broadcom and Marvell have spent the past two years building into a serious Nvidia alternative. Broadcom’s OpenAI partnership, announced October 13, 2025, covers 10 gigawatts of custom accelerators in a deal Reuters and the Wall Street Journal both described as worth roughly $10 billion, with initial chips (code-named Jalapeño) targeted for deployment by the end of 2026. Broadcom separately signed a long-term agreement with Google on April 6, 2026, covering custom Tensor Processing Units and AI rack components through 2031, and is co-developing Meta’s second-generation MTIA accelerator.

Marvell, meanwhile, has been AWS’s custom silicon partner since well before this Qualcomm deal existed. The two companies expanded their relationship in a five-year, multi-generational agreement announced December 2, 2024, with Marvell serving as lead design partner on AWS’s Trainium3 and Trainium4 chips. Marvell also works with Microsoft on the Maia 100 and Maia 200 AI accelerators and, as of an August 19, 2026 disclosure reported by The Register, added a second major custom chip agreement with Google.

Company	Hyperscaler Partner	Deal Announced	Reported Value / Scope
Qualcomm	Meta, Microsoft (Dragonfly C1000/AI300)	June 24, 2026	Not publicly disclosed
Qualcomm	Amazon Web Services	September 8, 2026	Up to $60B ceiling; ~$4B warrant
Broadcom	OpenAI	October 13, 2025	~$10B; 10 gigawatts of accelerators
Broadcom	Google	April 6, 2026	Custom TPUs and rack components through 2031
Marvell	Amazon Web Services	December 2, 2024 (expanded)	5-year, multi-generational; Trainium3/4 design
Marvell	Microsoft	Ongoing (Maia 100/200)	Not publicly disclosed

Set against that backdrop, Qualcomm’s entry looks less like a first mover and more like a fast follower trying to claim a seat at a table Broadcom and Marvell already occupy at every major hyperscaler. What Qualcomm brings that neither competitor can fully match is a low-power engineering pedigree built over two decades of smartphone chip design, which is precisely the skill set inference workloads reward.

How Big Is the Custom AI Silicon Market, Really

The dollar figures attached to individual deals only make sense against the size of the market they’re competing for. Estimates vary sharply depending on what analysts count as “custom AI silicon,” but the direction is consistent: this is one of the fastest-growing segments in semiconductors. Global Market Insights puts the ASIC segment of the AI accelerator chips market at roughly $35.6 billion in 2026, inside a broader $154.6 billion total accelerator market. J.P. Morgan’s research, by contrast, frames the custom AI ASIC opportunity more broadly at $60 billion to $70 billion in 2026. Tom’s Hardware’s May 2026 analysis of the sector found that custom ASIC shipments are projected to grow 44.6% year-over-year in 2026, nearly triple the 16.1% growth rate forecast for merchant GPUs.

Whichever figure ends up closest to reality, the trend line explains why Qualcomm wanted in. A market growing at more than 40% a year, even from a smaller base than Nvidia’s GPU business, is worth the multi-year engineering investment required to design custom inference silicon for a single customer’s specific workloads.

The Optical Connectivity Piece Nobody’s Talking About Enough

Most coverage of the Qualcomm AWS deal has focused on the chip side, but the 1.6 Tb/s optical connectivity component deserves more attention than it’s getting. As AI clusters scale past tens of thousands of accelerators, the interconnect between chips, not the chips themselves, increasingly determines how fast a training run or inference batch actually completes. Nvidia has spent heavily on its own NVLink and Spectrum-X networking to keep that bottleneck under control, and Broadcom’s custom silicon business already includes substantial networking and optical DSP revenue alongside its accelerator work.

Qualcomm entering optical connectivity alongside inference chips suggests the company is trying to sell AWS a more complete rack-level solution rather than a single component that has to be integrated into someone else’s networking stack. That’s a more ambitious pitch than Qualcomm has made in the data center before, and it puts the company in more direct competition with Broadcom’s networking silicon business, not just its custom accelerator business.

Market Reaction and What Wall Street Is Watching

Qualcomm shares moved sharply higher on announcement day, with Tech Times and multiple financial outlets reporting an intraday gain as high as 7%. That reaction reflects how starved investors have been for evidence that Qualcomm’s data center pivot, first signaled with Dragonfly C1000 in June, has real hyperscaler demand behind it rather than just a roadmap slide. A confirmed AWS relationship, even one structured around a purchasing ceiling rather than a guaranteed order book, gives analysts a second data point after Meta and Microsoft to argue Qualcomm’s server ambitions are more than a smartphone company chasing a trend.

The warrant structure itself is also a signal worth reading carefully. AWS didn’t pay cash for exclusive access to Qualcomm silicon. It received equity upside tied to its own future purchasing behavior, which means Amazon’s incentive to actually buy at scale is now partly aligned with Qualcomm’s stock price. That’s a financing structure increasingly common in AI infrastructure deals throughout 2025 and 2026, where hyperscalers use their purchasing power as leverage to extract equity stakes rather than simply negotiating on price.

What This Means for Nvidia

No single custom silicon deal threatens Nvidia’s near-term dominance. Even generous estimates of the custom ASIC market, in the $35 billion to $70 billion range for 2026, remain a fraction of the broader AI accelerator market that multiple analyst firms peg above $150 billion this year, with Nvidia still capturing the large majority of that revenue. But the cumulative effect of Broadcom’s OpenAI and Google deals, Marvell’s AWS and Microsoft relationships, and now Qualcomm’s Meta, Microsoft, and AWS agreements adds up to a hyperscaler landscape where every major cloud provider now has at least one, and often two, non-Nvidia custom silicon partners actively shipping or developing inference chips.

That doesn’t replace Nvidia GPUs for training frontier models, where CUDA software lock-in remains a real moat. It does chip away at the inference side of the business, where workloads are more predictable, more cost-sensitive, and easier to optimize for with purpose-built silicon. Nvidia’s own market share estimates for 2026, already down from a 2024 peak near 87% to a range most analysts place between 75% and 85%, reflect that erosion is already underway, deal by deal, well before Qualcomm entered the picture.

Historical Context: Qualcomm’s First Data Center Attempt

It’s worth remembering why Qualcomm left the server business in the first place. The Centriq line, launched with real hyperscaler interest in 2017, never found a sustainable customer base against Intel’s then-dominant Xeon franchise, and Qualcomm shut the project down in 2018 rather than keep funding a business with no clear path to profitability. The market Qualcomm is reentering in 2026 looks nothing like the one it left. AI inference demand didn’t exist as a distinct, massive workload category in 2018, and the willingness of hyperscalers to fund custom silicon partners, rather than insisting on off-the-shelf general-purpose chips, has completely reshaped what a viable data center chip business looks like.

That shift, from general-purpose CPUs competing head-on with Intel to purpose-built inference accelerators co-designed with a specific hyperscaler, is the reason Qualcomm’s second attempt has a real shot where the first one failed.

Predictions: Where the Qualcomm AWS Deal Goes From Here

Based on the deal structure, the pattern Qualcomm set with Dragonfly C1000, and the broader trajectory of custom AI silicon through 2026, a few outcomes look likely over the next 12 to 24 months.

Qualcomm will likely name its fourth hyperscaler customer within the next two quarters, following the pattern of sequential announcements set by Meta, Microsoft, and now AWS.
Initial AWS deployment of Qualcomm inference silicon probably won’t reach production scale before late 2027, given the multi-year design cycles typical of custom hyperscaler chips and the ten-year warrant horizon built into the deal.
The $60 billion ceiling will almost certainly not be reached on a fixed schedule, and expect periodic analyst debate over how much of it AWS actually intends to spend versus how much exists purely to structure the warrant.
Qualcomm’s optical connectivity ambitions will draw more competitive attention from Broadcom, which has far more entrenched networking silicon relationships across the hyperscaler landscape.
Nvidia’s data center accelerator revenue share will keep drifting lower through 2027 as inference workloads increasingly shift toward custom silicon, even as Nvidia remains the dominant choice for frontier model training.

Risks: Why the $60 Billion Number Deserves Skepticism

Every outlet that covered this deal carefully flagged the same caveat: $60 billion is a ceiling, not a commitment. AWS is under no contractual obligation to spend anywhere near that figure, and the warrant mechanics mean Qualcomm only benefits from the full equity upside if Amazon actually follows through with purchases over a full decade. Custom silicon programs also carry execution risk that pure chip-supply deals don’t. Qualcomm has to design, validate, and ship inference silicon that meets AWS’s specific performance and power targets, a process that has derailed or delayed other hyperscaler custom chip programs in the past.

There’s also the question of whether AWS genuinely needs a third external chip supplier on top of its own Trainium roadmap and its existing Nvidia GPU fleet. Marvell’s Amazon relationship already covers a meaningful share of AWS’s custom silicon design work. Qualcomm will need to demonstrate a performance-per-watt or cost advantage specific enough to justify AWS running three parallel chip strategies at once, rather than simply doubling down on Trainium.

What It Means for Enterprise AI Buyers and Developers

For companies building on AWS today, the immediate practical impact is close to zero. No Qualcomm inference chip is shipping in an AWS region yet, and the deal’s own ten-year vesting horizon signals this is infrastructure planning for the second half of the decade, not a service available next quarter. The more relevant near-term signal for developers and IT buyers is what this deal says about where inference pricing is headed. More competition among chip suppliers serving the same hyperscaler, Qualcomm alongside Marvell-designed Trainium and Nvidia GPUs, historically pushes cloud providers toward more aggressive inference pricing as they diversify away from any single supplier’s cost structure and margin requirements.

Enterprises evaluating multi-year AI infrastructure commitments should treat this as one more data point confirming that the inference hardware layer is going to keep fragmenting across specialized chips rather than consolidating around a single vendor, a trend that argues for keeping inference workloads portable across hardware backends wherever practical.

Frequently Asked Questions

What exactly did Qualcomm and AWS announce on September 8, 2026?

A multi-year, multi-generational partnership to co-develop custom AI inference chips and optical connectivity hardware for AWS data centers, structured around a warrant giving Amazon the right to buy up to 25 million Qualcomm shares tied to as much as $60 billion in future purchases, according to Qualcomm’s 8-K filing reported by Futurum Group and Tech Times.

Is the $60 billion a guaranteed order for Qualcomm?

No. Multiple outlets, including The Register and TechTarget, describe it explicitly as a ceiling that determines how much of the stock warrant vests, not a binding purchase commitment from Amazon.

When will AWS actually use Qualcomm’s AI chips?

No public shipping timeline has been disclosed. Given the ten-year warrant expiration of September 3, 2036, and typical multi-year custom silicon design cycles, production deployment is unlikely before 2027 at the earliest.

How does this compare to Qualcomm’s Meta and Microsoft deals?

Qualcomm named Meta and Microsoft as customers for its Dragonfly C1000 server CPU and AI300 inference accelerator at its June 24, 2026 Investor Day, and said at the time it was working with two more unnamed hyperscalers. AWS is now confirmed as one of those additional partners.

Does this replace AWS’s in-house Trainium chips?

No. AWS continues developing its own Trainium and Inferentia silicon with Marvell as lead design partner under a five-year agreement expanded in December 2024. The Qualcomm deal adds a third chip strategy alongside in-house Trainium and third-party Nvidia GPUs rather than replacing either.

How did Qualcomm’s stock react to the announcement?

Qualcomm shares gained as much as 7% intraday on September 8, 2026, according to Tech Times, reflecting investor confidence that the company’s data center pivot has real hyperscaler demand behind it.

Who are Qualcomm’s main rivals in custom AI silicon?

Broadcom and Marvell are the two most established players. Broadcom holds a roughly $10 billion, 10-gigawatt custom accelerator deal with OpenAI and a long-term Google TPU agreement through 2031. Marvell has been AWS’s lead custom silicon design partner since a five-year deal expanded in December 2024 and also works with Microsoft on the Maia accelerator line.