Apple has held talks with Nvidia about licensing NVLink Fusion, the chip-to-chip interconnect technology that Nvidia opened to outside silicon makers last year, according to a report from The Information that was corroborated on September 16 and 17, 2026 by Reuters, MacDailyNews, 9to5Mac, Tech Times, Bloomberg, and the Hong Kong-based outlet Futu News. The story is being framed around Apple’s tentative return to server hardware, its first since it killed the Xserve line in 2011. But the detail that matters more for the semiconductor and cloud industries is narrower and more interesting: Apple, a company that has spent fifteen years building every piece of its own silicon and software stack in-house, is reportedly willing to plug into Nvidia’s proprietary networking fabric to make its AI hardware work at data-center scale.
That single decision, if it holds, says more about the state of AI infrastructure in September 2026 than it does about Apple’s product roadmap. Even a company with Apple’s balance sheet and its own interconnect experience (Apple already ships proprietary fabrics inside the Mac Pro and its M-series Ultra chips) apparently sees no path to competitive AI server networking without Nvidia’s help. This piece looks at what is actually confirmed, what Nvidia’s NVLink Fusion does, why Apple would need it, and what it means for the wider AI infrastructure market, including AWS, Microsoft Azure, Google Cloud, Broadcom, and Marvell.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What The Information’s report actually says
According to the original report, cited by MacDailyNews and 9to5Mac, Apple is developing an enterprise AI inference server built around its still-unreleased M8 Ultra chip. Two configurations are reportedly under consideration: a smaller unit clustering two M8 Ultra processors and a larger one with four. The target buyers are described as AI developers, enterprises, and government agencies that want to run models on hardware they own rather than rent capacity from a public cloud provider.
The plans are explicitly not finalized. Apple’s own engineers reportedly view Nvidia’s NVLink Fusion as the best available option for linking multiple M8 Ultra chips inside a rack, but the project could still be shelved, delayed, or shipped with a different interconnect entirely. Tech Times put the likely timeline at no earlier than 2029, which places this firmly in the “early-stage internal project” category rather than an imminent product launch. Neither Apple nor Nvidia has issued an on-record confirmation; every detail traces back to unnamed sources described by The Information as familiar with the discussions.
Still, the framing matters. This is not a rumor about a new iPhone chip or a software feature. It is a claim that Apple is exploring commercial data-center hardware for the first time in fifteen years, and that it cannot do so on Apple silicon alone. That is a meaningfully different story than most Apple product leaks, and it explains why outlets that don’t normally cover enterprise infrastructure, including consumer-focused Mac blogs, picked it up within hours.
Why Apple needs Nvidia’s networking, not just its chips
Apple’s M-series Ultra chips already use a proprietary interconnect called UltraFusion to bond two Max dies into a single package, so Apple clearly knows how to build high-bandwidth links between its own silicon. The problem NVLink Fusion solves is different: it is about connecting many chips, potentially dozens, across multiple servers inside a single AI cluster, not fusing two dies on one package.
Nvidia opened NVLink Fusion to third-party silicon vendors at Computex in Taipei in May 2025. Nvidia CEO Jensen Huang described the goal directly at that keynote: “NVLink Fusion opens Nvidia’s AI platform and rich ecosystem for partners to build specialized AI infrastructures,” he said, according to the Taipei Times. Huang also framed the offering as something broader than a chip license. “NV link fusion is so that you can build semi-custom AI infrastructure, not just semi-custom chips,” he told reporters, a remark reported by CNBC at the time. By August 2026, Nvidia had widened that licensing scope to cover two distinct categories of outside silicon, custom XPUs and third-party CPUs, rather than limiting the program to accelerator chiplets alone, underscoring how quickly the ecosystem Huang described has broadened in the fifteen months since launch.
That distinction is the crux of the Apple story. Nvidia is not just selling switch chips. It is selling admission into an entire rack-scale ecosystem, the cabling standards, the switch silicon, the software stack, and the coherency protocol that lets custom accelerators from outside vendors sit inside a fabric otherwise built around Nvidia’s own GPUs. The current generation, NVLink 6, can connect up to 72 accelerators in a single non-blocking domain at roughly 3.6 TB/s per accelerator, an all-to-all topology that is difficult to replicate from scratch. For a company entering enterprise AI hardware for the first time since Xserve, building an equivalent fabric independently, complete with a competitive software ecosystem persuading customers to adopt it, would likely take years Apple does not have if it wants any credible AI infrastructure position before the end of the decade.
Xserve, and why Apple walked away from servers once already
Apple’s server history is short and not particularly happy. Xserve launched in 2002 as a 1U rack-mount Power Mac aimed at web hosting, media production, and academic computing. It never captured meaningful market share against Dell, HP, and IBM, and Apple discontinued it in January 2011, redirecting the company fully toward consumer devices and, later, services. For a decade and a half, Apple’s position on enterprise infrastructure was essentially: we make the client, someone else makes the server.
What changed is the economics of AI inference. Running large models locally, for privacy, latency, cost, or regulatory reasons, has become a genuine enterprise and government requirement, and Apple’s M-series chips already post competitive performance-per-watt numbers in the workstation space via the Mac Studio and Mac Pro. The strategic logic reported by The Information is that Apple already has the silicon efficiency; what it lacks is the data-center-grade networking and software ecosystem to sell that silicon as a rack-scale product to businesses, rather than as a desktop under someone’s desk. If Apple is serious about a Mac Studio-derived server business, it is a bet that on-premise AI inference demand from privacy-sensitive customers, healthcare systems, defense contractors, financial firms, and sovereign-cloud governments, is now large enough to justify rebuilding a business Apple abandoned once already.
The AI infrastructure market Apple would be entering
Apple would not be entering an empty field. Amazon Web Services, Microsoft Azure, and Google Cloud dominate cloud AI compute, and all three now also design their own AI accelerators (Trainium, Maia, and TPU respectively) specifically to reduce dependence on Nvidia GPUs. Nvidia itself, meanwhile, sells complete rack-scale AI systems under its own brand, competing at the systems level with the very companies it licenses NVLink Fusion to. That overlap, Nvidia as both networking supplier and systems competitor, is one reason analysts have questioned how enthusiastically Nvidia will support a partner like Apple that could eventually sell a competing enterprise inference box.
Qualcomm has taken a similar path recently, striking a large AI infrastructure supply arrangement with Humain, the Saudi Arabia-backed AI company, this year, a sign that even chip designers with deep systems experience are choosing partnership over building rival rack-scale ecosystems from zero. Apple weighing the same trade-off, buying access to Nvidia’s fabric rather than building a competing one, fits that broader industry pattern rather than breaking from it.
| Company | AI infrastructure approach | Interconnect strategy | Primary customer |
|---|---|---|---|
| Apple (reported) | M8 Ultra enterprise inference server, 2x or 4x chip configurations | Reportedly in talks to license Nvidia NVLink Fusion | Enterprises, developers, governments wanting on-premise AI |
| Nvidia | Full rack-scale AI systems (GB-series) plus licensed NVLink Fusion fabric | Proprietary NVLink 6, opened to third-party silicon since May 2025 | Hyperscalers, enterprises, AI labs |
| AWS | Trainium custom accelerators plus Nvidia GPU instances | Proprietary EFA networking; also buys Nvidia NVLink systems | Cloud tenants of all sizes |
| Microsoft Azure | Maia custom accelerators plus Nvidia GPU instances | Proprietary fabric; also buys Nvidia NVLink systems | Enterprise and OpenAI workloads |
| Google Cloud | TPU custom accelerators plus Nvidia GPU instances | Proprietary Inter-Chip Interconnect (ICI); also Nvidia NVLink | Enterprise and internal AI workloads |
| Qualcomm | AI200/AI250 accelerators under a reported multi-year supply deal with Humain (Saudi Arabia) | Partnering with hyperscaler infrastructure rather than building rival fabric | Humain (Saudi Arabia) data centers |
NVLink Fusion, explained for engineers who haven’t touched it
NVLink started life in 2016 as a way to connect Nvidia GPUs to each other faster than PCIe allowed. NVLink Fusion, announced at Computex 2025, is a licensing program that opens pieces of that stack, switch silicon, chiplet IP, and software, to outside CPU and accelerator vendors, so their chips can sit inside an Nvidia-designed rack alongside Nvidia GPUs, or in Apple’s reported case, alongside each other using Nvidia’s switching layer. Reuters, in a report on the original announcement summarized by MarketScreener, described it plainly: Nvidia would “sell tech to speed AI chip communication” to outside chip designers building custom multi-chip AI systems. As of August 2026, Reuters further described the licensed NVLink Fusion chiplet itself as the mechanism that stitches outside custom XPUs into a single rack-scale fabric, with Nvidia characterizing the whole offering as a bundled package of switches, chiplet designs, and software rather than a standalone networking chip.
For a developer or infrastructure engineer, the practical implication is straightforward. Instead of writing custom RDMA or proprietary fabric code to scale an inference workload across chips from a non-Nvidia vendor, NVLink Fusion licensees can lean on the same low-latency, high-bandwidth primitives that CUDA-based software already assumes. That lowers the software burden for anyone building on top of the resulting hardware, including, eventually, developers targeting Apple’s rumored server if it ships. It also means Apple’s inference stack, likely built on its Metal and Core ML frameworks today, would need a bridging layer capable of speaking Nvidia’s interconnect protocol, a nontrivial software undertaking on top of the hardware licensing question.
Why now: the on-premise AI inference market is turning real
The timing lines up with a broader shift reported across the industry through 2026: enterprises and governments increasingly want AI inference capacity they physically control, rather than API calls to a third-party cloud model. Regulated industries, healthcare, defense, and financial services chief among them, have data residency and auditability requirements that public cloud AI services don’t always satisfy cleanly. Sovereign AI, the idea that a country wants its own AI compute capacity rather than dependence on US hyperscalers, has become a policy talking point in Europe and Asia through 2026 as well.
Apple, notably, is positioned differently from a typical enterprise hardware vendor here. Its brand already carries a privacy-first reputation from a decade of on-device processing messaging around Apple Intelligence. An enterprise AI box that runs inference locally, backed by that same privacy positioning, is a plausible pitch to exactly the customers who are most wary of sending proprietary data to a shared cloud model. Whether Apple can execute the enterprise sales motion, support contracts, hardware lifecycle management, that this market demands is a separate and much harder question than whether the silicon is fast enough.
The M8 Ultra: what’s actually known
Apple has not announced the M8 Ultra, and no official specifications exist as of this writing. What is reported is limited to its role in this server project: it would be Apple’s highest-performance processor at the time of the server’s hypothetical launch, offered in two-chip and four-chip clustered configurations inside the enterprise box. Apple’s naming convention suggests M8 Ultra would follow the current M-series Ultra pattern, two Max-class dies fused via UltraFusion into one package, the same approach used in the M3 Ultra and its predecessors, scaled up for a future process node. Given the reported 2029 target, M8 Ultra sits several silicon generations beyond Apple’s current shipping M5 family, and any specific performance or efficiency numbers attributed to it today should be treated as speculative until Apple confirms them directly.
Market reaction and analyst framing
Because the report rests largely on anonymous sourcing, market reaction so far has been measured rather than dramatic, even after Reuters independently reported in September 2026 that Apple had held talks with Nvidia about NVLink Fusion for M8 Ultra-based servers targeted around 2029. Neither Apple nor Nvidia stock moved sharply on the news, consistent with a story that describes an exploratory, multi-year-out initiative rather than a signed supply agreement or product announcement, and Reuters was careful to note the plan remains unconfirmed and could still be scrapped. Coverage from financial outlets including Bloomberg and Investing.com framed the story primarily as evidence that Apple sees genuine long-term upside in enterprise AI hardware, not as an immediate competitive threat to existing server vendors like Dell, HPE, or Supermicro.
The more interesting read for industry watchers is what the talks say about Nvidia’s negotiating position. Futu News, an outlet focused on Hong Kong and mainland Chinese markets, summarized the dynamic directly: “Apple is developing an enterprise-grade server for AI applications, planned to be equipped with its self-developed M8 Ultra chip, and has already held talks with NVIDIA about adopting its NVLink Fusion interconnect technology,” according to a Futu News report citing The Information. Even a company as fiercely vertically integrated as Apple, having built its own CPUs, GPUs, neural engines, and modems in-house over the past decade specifically to reduce supplier dependence, is reportedly choosing to license rather than build when it comes to AI networking. That is a meaningful data point about how defensible Nvidia’s networking moat has become, independent of GPU market share.
What this means for Broadcom, Marvell, and rival interconnect vendors
Broadcom and Marvell have spent the past two years pitching custom AI accelerator design services and Ethernet-based scale-out networking (via the Ultra Ethernet Consortium) as an open alternative to Nvidia’s proprietary NVLink stack. If Apple, a company with essentially unlimited engineering resources and its own chip design history, is nonetheless leaning toward NVLink Fusion rather than an Ethernet-based or custom fabric, that is a reputational point in Nvidia’s favor in the broader interconnect debate. It does not settle the argument, Ethernet-based scale-out remains cheaper and more open for hyperscalers running mixed fleets, but it undercuts the narrative that only smaller or less sophisticated chip designers need Nvidia’s fabric. Apple reportedly evaluating and leaning toward NVLink Fusion despite having every resource to build its own alternative is a stronger signal for the interconnect market than if a smaller AI chip startup had made the same choice.
| Interconnect approach | Backing | Max scale (reported) | Openness | Primary use case |
|---|---|---|---|---|
| NVLink 6 / NVLink Fusion | Nvidia | Up to 72 accelerators per domain, ~3.6 TB/s per XPU | Licensed to approved third parties since May 2025 | Rack-scale AI training and inference clusters |
| UltraFusion | Apple | 2 dies per package (current M-series Ultra) | Apple-internal only | Single-package CPU/GPU fusion, not rack networking |
| Ultra Ethernet Consortium (UEC) | Broadcom, Arista, AMD, Meta, others | Scales via standard Ethernet switching | Open standard | Hyperscaler scale-out networking |
| Proprietary hyperscaler fabrics (EFA, ICI, etc.) | AWS, Google | Varies by generation, internal only | Closed, vendor-specific | Internal custom accelerator clusters |
Historical context: Apple’s on-again, off-again enterprise ambitions
Apple has flirted with enterprise infrastructure repeatedly since Xserve’s 2011 exit, always at arm’s length. Mac Pro and Mac Studio units are common in rendering farms and media production pipelines, but Apple has never marketed them as data-center products with the support contracts and rack-density design that enterprise buyers expect. Apple’s own cloud infrastructure, including Apple Intelligence’s Private Cloud Compute servers introduced in 2024, already runs on Apple silicon at scale internally, which is itself a meaningful precedent: Apple has proven it can operate M-series chips as server-class inference hardware for its own workloads. What has always been missing is a commercial product built for outside customers, exactly the gap this reported M8 Ultra server would attempt to fill.
What could go wrong for Apple
Several risks separate this reported project from a shipped product. First, the reported 2029 timeline means Apple would be entering enterprise AI inference years after competitors have established relationships with the exact customers Apple wants, healthcare systems, defense contractors, and governments typically run multi-year procurement cycles and vendor lock-in that make late entry costly. Second, licensing NVLink Fusion means depending on a direct competitor, Nvidia’s own DGX and rack-scale systems compete for the same enterprise inference budget Apple would be chasing, an awkward position for any long-term partnership. Third, Apple has no enterprise sales, support, or systems-integration organization comparable to Dell, HPE, or Cisco, building or acquiring one is a multi-year undertaking independent of the hardware itself. Any of these factors could push the project past 2029 or kill it outright, exactly the caveat every outlet covering the story has attached to it.
Five predictions for how this plays out
- Apple will not confirm or deny the report directly; expect silence through at least its next two product launch cycles, consistent with its standard practice on unreleased hardware.
- If the NVLink Fusion licensing deal closes, expect it to be disclosed indirectly, through a regulatory filing, a Nvidia developer conference mention, or a supply-chain report, rather than a joint press release.
- The 2029 timeline will likely slip further right, in line with how most multi-year Apple silicon roadmap leaks have historically moved compared to initial reports.
- Even if the server ships, expect Apple to target a narrow enterprise niche (privacy-sensitive on-premise inference) rather than compete broadly against AWS, Azure, and Google Cloud on price or scale.
- Regardless of Apple’s outcome, expect more non-hyperscaler chip designers to publicly evaluate NVLink Fusion licensing over the next 12 months, reinforcing Nvidia’s networking position even among companies that compete with it elsewhere.
How this compares to Apple’s other recent infrastructure moves
This report doesn’t exist in isolation. Apple has spent 2026 signaling a broader willingness to work with outside silicon partners where its own stack falls short, a shift from the fully vertically integrated posture it held for most of the past decade. The pattern across Apple’s 2026 infrastructure and roadmap reporting, from smart home hub plans extending to 2028 to this server initiative, suggests a company increasingly comfortable admitting where partnerships beat in-house builds, at least for infrastructure that customers never see or touch directly, as opposed to consumer-facing silicon like the A-series and M-series chips Apple continues to design entirely itself.
What developers and IT buyers should actually do with this news
For most engineering teams, the practical answer is: nothing, yet. A 2029-earliest timeline for an unconfirmed product means procurement teams evaluating on-premise AI inference hardware today should continue comparing options that already ship, Nvidia DGX systems, AMD Instinct-based servers, and the growing field of Arm-based inference boxes, rather than waiting on a speculative Apple product years out. Where this story is useful is as a signal of direction: on-premise, privacy-preserving AI inference is a large enough market that even Apple, historically allergic to enterprise hardware, is reportedly willing to reconsider it. Teams building AI infrastructure strategy for the next three to five years should treat that direction, not Apple’s specific product, as the actionable takeaway.
Frequently asked questions
Is Apple actually building an AI server?
Apple has not confirmed this. The claim comes from a report by The Information, corroborated by outlets including Reuters, MacDailyNews, 9to5Mac, Tech Times, and Bloomberg, all citing anonymous sources, with Reuters’ own September 2026 reporting likewise stressing that Apple’s plans are not finalized and could still change or be canceled.
What is NVLink Fusion?
NVLink Fusion is a licensing program Nvidia launched at Computex 2025 that lets third-party chip designers connect their own CPUs and accelerators using Nvidia’s high-bandwidth NVLink interconnect and switch technology, instead of building a competing fabric from scratch.
What is the M8 Ultra chip?
The M8 Ultra is an unreleased, unannounced Apple silicon chip reportedly planned as the processor for this enterprise AI server. No official Apple specifications exist yet; it would sit several generations beyond Apple’s current M5 family.
When would Apple’s AI server launch?
Reports, including from Tech Times, put the earliest plausible launch around 2029, and stress the project could still be delayed or canceled entirely before then.
Why does Apple need Nvidia if it makes its own chips?
Apple’s own interconnect technology, UltraFusion, links two dies inside a single chip package. It was not designed to network many chips together across a data-center rack at the scale AI inference clusters require, which is the specific problem NVLink Fusion addresses.
Did Apple make servers before?
Yes. Apple sold the Xserve rack-mount server line from 2002 until discontinuing it in January 2011, after which it exited the dedicated server hardware business entirely until this reported project.
Who would buy Apple’s AI server?
Reports describe the target customers as enterprises, AI developers, and government agencies seeking on-premise AI inference capacity rather than public cloud AI services, likely for privacy, latency, or regulatory reasons.
Does this compete with AWS, Azure, and Google Cloud?
Not directly, based on current reporting. The described product targets on-premise, privately-owned AI inference hardware rather than public cloud rental capacity, positioning it closer to enterprise server vendors like Dell and HPE than to hyperscale cloud providers.
