Everything About AI
NextFin News - Meta and Apple, two Silicon Valley rivals that spent years trading insults over privacy and the future of computing, announced on Thursday a joint effort to run one of the most powerful open-weight AI models ever released directly on Apple's high-end Mac computers. The collaboration is a signal that the AI race is redrawing the map of industry alliances - and that the next bottleneck in artificial intelligence may be memory bandwidth, not cloud scale.
Meta released Muse Spark Open, which it calls the most capable U.S.-developed open-weight model available. The model has 816 billion total parameters, of which only 42 billion are active at any given time, and a one-million-token context window. To prove it can run on a desktop, Meta worked with Apple to optimize the model for the Mac Studio powered by the M5 Ultra chip, using Apple's latest version of MLX, an open-source array framework built for Apple silicon's unified memory architecture. On a cluster of four Mac Studios, Apple says the model generates more than 36 tokens per second - faster than average human reading speed.
The market barely blinked. Meta closed at $736.72, down 0.11 percent, with a market capitalization of about $1.88 trillion. Apple closed at $342.00, down 0.28 percent, at a market value of about $4.99 trillion. The muted reaction came after a volatile week: Apple's stock pushed its market value above $5 trillion for the first time on Wednesday, while Meta shares fell 3.27 percent that day. Meta's stock has still risen about 30 percent over the past three months, buoyed by a string of AI launches.
The real story is not the day's tape. It is that the two companies who once defined the privacy wars are now sharing a stack - and that a frontier-class model can fit on a desk if the hardware is built around one pool of memory instead of many.
The Deal: An Engineering Collaboration, Not a Truce
This is not a product bundling agreement or a revenue-sharing arrangement. It is an engineering collaboration: Meta optimized its model to run on Apple's hardware, and Apple published the MLX tools and benchmarks that make it possible. That distinction matters, because it tells you what each side is actually buying.
Meta gets a showcase for its rebooted AI effort. Muse Spark Open is the first model from the Meta Superintelligence Lab, the unit CEO Mark Zuckerberg unveiled in June after years of criticism that the company's AI models did not live up to expectations. Zuckerberg hired a host of AI heavyweights, including Scale.AI founder Alexandr Wang, to lead the push. Releasing the model as open weights - downloadable from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models - is a bid to win developers who are wary of depending on a single vendor's API.
"Congrats to the team! Amazing to run a model of Muse Spark Open's caliber and scale this efficiently, right from your desk. Mac Studio keeps pushing the boundaries of on-device AI!" Apple CEO John Ternus said in a reply to Zuckerberg on X.
Apple gets something equally valuable: proof that its silicon can handle frontier workloads. The MLX post is a hardware-first argument. Unified memory bandwidth, not cloud scale, is the bottleneck for local inference - and Apple's M5 Ultra delivers 1.2 TB/s of it, up from 819 GB/s on the M3 Ultra. Because generation speed is memory-bandwidth bound, Apple says the M5 Ultra delivers 1.6 times faster tokens per second than the M3 Ultra. Time to first token was up to 4.6 times faster, thanks to the GPU Neural Accelerators built into the M5 generation.
The timing under new leadership is not incidental. Ternus became Apple's CEO on September 1, succeeding Tim Cook. He has spent his career in hardware engineering rather than services, and the partnership reads like a hardware executive's answer to the AI era: put more memory closer to the compute, and let the model come to the chip.
The history makes the optics striking. Apple and Meta regarded each other warily for years, and their relationship hit rock bottom after Apple introduced app-tracking transparency in 2021 - a privacy change Meta estimated would cut its revenue by about $10 billion in 2022. Zuckerberg called Cook's privacy stance "extremely glib." The two companies have not struck a public partnership since 2017, when Apple removed system-level Facebook integration from the iPhone. "We're excited for mommy and daddy to be playing nice again," Alexandr Wang, Meta's chief AI officer, said of the collaboration.
The Mechanism: How 816 Billion Parameters Fit on a Desk
The trick is not that the model is small. At 816 billion parameters, Muse Spark Open is larger than most open models. The trick is that it is a mixture of experts: only 42 billion parameters - about 5 percent - are active for any given token. That cuts the compute needed per step.
The second trick is Apple's unified memory. In a conventional PC or server, the CPU, GPU and accelerators each have their own memory, and data must be copied between them. Apple's M5 Ultra gives every component access to the same memory pool, with up to 512 GB per Mac Studio. Two machines together can hold the entire model - a terabyte of unified memory - and Meta describes them as behaving "like one supercomputer."
The third piece is MLX itself. Apple's open-source framework is built around the unified-memory model from the ground up, so model weights sit in one place and every accelerator reads them without copying. Meta's optimization work on top of MLX is what closes the gap.
Apple's benchmarks show the payoff. At an 8,192-token context, Muse Spark Open decoded at 15.74 tokens per second with pipeline parallelization across two M5 Ultra Mac Studios, 24.99 tokens per second with tensor parallelization across two machines, and 36.73 tokens per second across four. For comparison, the dense 128-billion-parameter Mistral Medium 3.5 reached 23.37 tokens per second across four machines, and the 27-billion-parameter Qwen 3.8 reached 50.89. Peak power consumption was about 300 watts for the M5 Ultra Mac Studio - a fraction of what a comparable GPU server draws.
Meta also shipped custom, auto-tunable Helion GPU kernels in SGLang for the cloud path, delivering 1.03 to 3.95 times kernel-level speedups on NVIDIA and AMD GPUs and a 3 percent end-to-end speedup on AMD's MI350. The model is available now through API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Weights are released under the Muse Model License with limited commercial terms.
The Second-Order Effect: Edge Inference Versus the Data Center
The first-order read is obvious: Apple sells more Mac Studios, Meta gets distribution for its model. The second-order effect is about the inference market itself, and it is the part the market is not pricing.
If frontier-class inference can run on a desk, the unit economics of AI change at the margin. A developer who can run an 816-billion-parameter model locally does not need to pay a cloud provider for every token. That does not eliminate cloud inference - training, fine-tuning and high-throughput serving still need data centers - but it carves out a class of workloads that can move to the edge: coding assistants that read an entire codebase, video tools that review hours of footage, and enterprise agents that must keep data on-premises.
That is a direct challenge to the assumption that AI demand will keep scaling GPU orders indefinitely. Nvidia has dominated AI chips because data centers need them. If a meaningful slice of inference migrates to Apple silicon and other edge hardware, the growth rate of data-center GPU demand could slow sooner than the consensus expects. The trade is not "AI wins or loses." It is "where does the inference happen, and who gets paid for it?"
The open-weight dimension reinforces the shift. Muse Spark Open joins a wave of Western open models - Nvidia's Nemotron updates, Reflection AI and Mistral releases last week - that give companies an alternative to the Chinese AI ecosystem and to closed models from OpenAI, Google and Anthropic. Open weights let companies run, audit and fine-tune models without depending on a single vendor's API. Meta is releasing weights, not source code, and the license carries limited commercial terms - but for many developers, downloadable weights are enough.
The Counter-Thesis: A Demo, Not a Displacement
The strongest case against reading too much into this announcement is simple: the hardware is expensive and the use case is narrow.
Two Mac Studios with M5 Ultra chips sit at the top of Apple's desktop lineup - the M5 Ultra starts at $5,499, and fully configured units run well into five figures - far beyond what most developers or companies will deploy at scale. The headline 36-tokens-per-second figure requires four machines. For most production workloads, cloud inference remains cheaper per token and far easier to operate. Apple's own AI strategy on the iPhone and in Siri relies on small, efficient on-device models, not 816-billion-parameter systems. The desktop demo proves a point about capability, not about cost.
There is also a question of whether the partnership changes anything fundamental about the Apple-Meta relationship. The companies still compete over the next generation of computing - glasses, agents, and the interface that sits between users and the internet. Meta is also pushing its own consumer hardware. This is a narrow engineering collaboration, not a truce.
The counter-thesis has force, but it mistakes the near term for the whole arc. Hardware costs fall. M-series chips have roughly doubled memory bandwidth every generation, and unified memory scales by adding machines rather than building custom servers. What looks like a niche demo today is a proof of concept for a class of inference that migrates to the edge as the economics improve. The question is not whether the desktop replaces the data center tomorrow; it is how much of the inference curve bends toward the edge over the next few years.
Cyclical or Structural: The Call
This is the judgment the announcement forces. Is the move of inference to the edge a cyclical fluctuation - a cost-driven wobble that will revert when cloud prices fall - or a structural shift that will not revert on its own?
The answer is both, and they must be separated. The cyclical leg is real: some workloads will move to the edge because cloud inference is expensive right now, and if cloud prices drop sharply, some of those workloads will move back. That is mean-reverting by nature.
But the structural leg is the one that matters. Three pieces of evidence point to a regime change rather than a wobble. First, the constraint is architectural, not just price: unified memory eliminates the data-copying bottleneck that has defined von Neumann machines for decades, and that does not revert. Second, privacy and data-sovereignty requirements - in healthcare, finance and enterprise software - create a permanent demand for models that never leave the premises. Third, the model architecture itself is moving toward efficiency: mixture-of-experts designs keep only a fraction of parameters active, which is exactly the shape of model that edge hardware can run. A cyclical claim would require three historical cycles of edge inference rising and falling back to the cloud; there is no such history, because the hardware did not exist. A structural claim requires evidence of a permanent regime change; the architecture, the privacy rules and the model design all supply it.
The market has not priced this because it is still thinking in the old frame: AI demand equals GPU orders. The gap between what is priced and what is happening is the opportunity - and the risk.
What to Watch
Meta reports earnings on October 21, 2026, and investors will be looking for evidence that AI spending is translating into revenue. The company expects to spend between $115 billion and $135 billion on AI infrastructure in 2026, up from $72.22 billion in 2025. Revenue is expected to reach $247.7 billion in 2026, up from $198.8 billion in 2025.
Three signals would confirm that this partnership is more than a demo: Apple publishes follow-on benchmarks showing frontier models running on fewer or cheaper machines; enterprise developers announce on-premises deployments of open-weight models on Apple-silicon clusters; or cloud inference pricing comes under pressure as edge alternatives gain share.
The falsifying signal is specific: if Meta and Apple do not announce a second joint optimization within twelve months, and if Mac Studio revenue shows no AI-related lift through fiscal 2027, the collaboration should be read as a one-off marketing exercise rather than a structural shift.
Conclusion: The Stack Is the Story
Short term, the announcement is a sentiment win for both companies and a validation of Meta's rebooted AI effort. Medium term, it gives Apple a showcase for M-series silicon in the AI era and gives developers a local option for frontier inference. Long term, the structural question is whether edge inference becomes a real substitute for cloud GPUs - and that depends on memory bandwidth, model efficiency and the price of silicon, not on a single launch.
Scenarios for the next twelve months: the base case is that Apple and Meta announce incremental optimizations and a handful of enterprise pilots, with Mac Studio demand firming but not transforming. The upside case is that memory bandwidth doubles again with the next generation and a major enterprise workload migrates on-premises, pulling the edge-inference narrative into the mainstream. The downside case is that the hardware cost stays prohibitive, cloud inference prices fall faster than edge economics improve, and the partnership fades into a footnote.
The collaboration does not end the Apple-Meta rivalry. It just shows that in the AI era, even enemies can share a stack when the economics demand it. The winners will not be the companies with the biggest models, but the ones that control where the model runs.
Data as of the October 15, 2026 market close. Prices from exchange data; all other figures from company and primary technical sources.
Explore more exclusive insights at nextfin.ai.
