Nvidia Buys Hugging Face to Lock Open-Source AI onto Proprietary GPUs

IBM’s mainframe chip collab, NVIDIA’s Poolside deal & Ox Alpha’s reveal — IBM

IBM

Nvidia is funding open-weight model ecosystems to ensure developers stay dependent on its proprietary CUDA silicon as frontier labs build custom in-house chips.

AI-assisted summary of IBM’s mainframe chip collab, NVIDIA’s Poolside deal & Ox Alpha’s reveal — original source. Check the original for full context.

Source published:

Key ideas

  • NVIDIA is rapidly consolidating open-source AI infrastructure through reported mega-deals, including a $12.9 billion acquisition of Hugging Face and a $6 billion transaction with code-generation startup Poolside. (0:00)
  • IBM revealed a dual-architecture mainframe processor developed with Arm at the Hot Chips conference, integrating IBM Z enterprise reliability with native Arm execution on a single silicon die. (13:14)
  • The stealth reasoning model 'OX Alpha' on OpenRouter, which claimed a throughput of 100 trillion tokens per day, was unmasked as the Chinese open-source model GLM 5.3. (24:34)

Detailed notes

## Background Published by IBM on August 28, 2026, as an RSS audio podcast, this 122nd episode of Mixture of Experts features hosts Tim Hwang and Aili McConnon alongside panelists Gabe Goodhart, Ash Minhas, and Skyler Speakman. According to the show description, the panel discusses IBM's Hot Chips conference reveal of a dual-architecture mainframe processor built with Arm to run Arm-native Linux workloads alongside z/OS, NVIDIA's massive dealmaking spree with Hugging Face and Poolside, and the unveiling of a mysterious model dubbed Ox Alpha.

## TL;DR The panel opens with NVIDIA's reported mega-acquisitions and investments in Hugging Face and Poolside, analyzing how the hardware giant uses open-source software and open model weights to drive demand for proprietary CUDA infrastructure while hedging against custom silicon. The discussion then moves to IBM's Hot Chips announcement of a dual-architecture chip combining IBM Z and Arm on a single die, examining how merging CISC and RISC execution allows mission-critical enterprises to run modern AI workloads directly alongside legacy data stores without sacrificing bulletproof reliability. Finally, the hosts dissect the stealth launch of Ox Alpha on OpenRouter—revealed right before recording to be China's GLM 5.3—and explore how masked deployments offer both unbiased technical testing and protection against aggressive release timing battles from frontier AI labs.

## Atmosphere & style The conversation carries the energetic, insider camaraderie of a panel of seasoned engineers and analysts geeking out over hardware silicon and AI strategy. Tim Hwang and Aili McConnon keep the dialogue brisk and journalistic, balancing Ash Minhas's infrastructure-level scrutiny with Gabe Goodhart's hands-on developer realism and Skyler Speakman's probing architectural skepticism. It feels conversational yet deeply technical, oscillating between lighthearted humor about shooting bullets through mainframe RAM sticks and sharp analysis of multi-billion-dollar corporate maneuvers in open-source AI.

## Bulletin - `00:00` NVIDIA is rapidly consolidating open-source AI infrastructure through reported mega-deals, including a $12.9 billion acquisition of Hugging Face and a $6 billion transaction with code-generation startup Poolside. - `13:14` IBM revealed a dual-architecture mainframe processor developed with Arm at the Hot Chips conference, integrating IBM Z enterprise reliability with native Arm execution on a single silicon die. - `24:34` The stealth reasoning model 'OX Alpha' on OpenRouter, which claimed a throughput of 100 trillion tokens per day, was unmasked as the Chinese open-source model GLM 5.3.

## Key takeaways - `00:00` NVIDIA's reported $12.9 billion acquisition of Hugging Face targets the central gravity of open-source AI, particularly the transformers reference library that defines open-weight architectures. - `00:00` An inherent tension exists between the open-source ethos of shrinking models for commodity silicon and NVIDIA's core business model of selling proprietary, high-end hardware. - `05:00` Open software and open weights still funnel compute back into proprietary hardware chokepoints, as third-party inference providers predominantly run on NVIDIA CUDA cores. - `05:00` Consolidating open-source model ecosystems serves as a defensive hedge for NVIDIA against frontier labs developing custom in-house ASICs and accelerators. - `10:00` NVIDIA aims to stimulate the entire model ecosystem ('letting a thousand flowers bloom') rather than climbing high up the stack to build walled consumer applications that compete with ChatGPT. - `13:14` IBM's joint dual-architecture processor with Arm allows enterprises to execute native Arm AI inference directly alongside mission-critical z/OS data on the same silicon, switching in nanoseconds. - `15:00` Combining Arm's RISC and mainframe CISC execution on one die resolves enterprise data transfer bottlenecks while raising questions about running fast-moving code in regulated core banking environments. - `18:40` The enduring value of IBM Z mainframes is extreme hardware fault tolerance, which enterprise clients cannot replace simply due to legacy COBOL maintenance challenges. - `20:00` Integrating Arm natively eliminates the multi-year cross-compilation struggle required to port modern Python and C runtimes like PyTorch onto isolated enterprise architectures. - `20:00` Mainframes will likely persist into the distant future because distributed cloud architectures cannot match their millisecond-level transactional latency. - `24:34` The mystery model OX Alpha stirred intense debate over its 100-trillion-token daily throughput claim before being revealed as the open-source Chinese release GLM 5.3. - `27:31` Stealth releases allow AI developers to sidestep release-day timing conflicts from competing frontier labs while mitigating regional and geographic testing biases.

## Detailed walkthrough ### `00:00` NVIDIA's Big Bets on Open AI: Hugging Face and Poolside

Tim Hwang opens episode 122 of *Mixture of Experts* by introducing co-host Aili McConnon alongside panelists Gabe Goodhart, Ash Minhas, and Skyler Speakman. While the panel plans to cover IBM's new next-generation dual-architecture mainframe processor revealed at the Hot Chips conference and the mystery behind the newly revealed Ox Alpha model, the immediate shockwave dominating the discussion is NVIDIA's aggressive dealmaking spree. Tim highlights reports of Hugging Face announcing an acquisition by NVIDIA for $12.9 billion, coming on the heels of a massive $6 billion transaction with AI code-generation startup Poolside. Gabe Goodhart flags these sweeping moves as being fundamentally about open-source AI, explaining that NVIDIA's core business relies on a sprawling, decentralized ecosystem rather than a monolithic closed provider.

Gabe lays out how NVIDIA has already pursued this path through its family of Nemotron models, pointing to the Poolside transaction as an engine to power that model line. Acquiring Hugging Face, however, is a play for the absolute center of gravity in open-source AI. Hugging Face is not just a hub for open-weight model hosting; its `transformers` Python library serves as the de facto reference implementation for nearly every open-source large language model in existence. Gabe emphasizes the practical reality for developers: if someone submits an architectural feature pull request to a downstream project like `llama.cpp` before it lands in Hugging Face's `transformers`, maintainers will reject or stall it until the upstream Hugging Face PR is merged.

At the same time, Gabe weighs the trade-offs of this move. Having the most well-capitalized hardware juggernaut backing the open ecosystem guarantees years of funding and stability for open-source practitioners. The potential conflict, however, lies in incentives: the open-source community inherently strives to make models smaller, cheaper, and capable of running across heterogeneous edge hardware and commodity silicon. Putting the ecosystem under NVIDIA—whose sole economic interest is selling massive, expensive, proprietary hardware—introduces an obvious tension. Drawing parallels to Microsoft's acquisition of GitHub, Gabe points out that Microsoft largely respected a firewall around GitHub’s open ethos while monetizing developer workflows through products like GitHub Copilot. The prevailing hope is that NVIDIA takes a similar stance with Hugging Face.

```mermaid flowchart TD accTitle: NVIDIA Open-Source Ecosystem Strategy accDescr: Flowchart illustrating how open-source software and open-weight models ultimately drive hardware monetization through NVIDIA CUDA cores.

subgraph OpenLayer["Open AI Ecosystem"] direction TB HF["Hugging Face Platform<br/>(Acquisition target: $12.9B)"] TR["Transformers Library<br/>(Upstream Reference Architecture)"] MOD["Open Weights & Models<br/>(Nemotron / Poolside Factory)"] HF --> TR TR --> MOD end

subgraph HardwareLayer["Proprietary Infrastructure Layer"] direction TB INF["Third-Party Inference Providers<br/>& Data Centers"] CUDA["NVIDIA CUDA Cores & GPUs<br/>(Hardware Lock-In)"] MOD --> INF INF --> CUDA end ```

- Tim Hwang frames the week's news as feeling like an "overwhelming amount of news," from IBM's dual processor to Ox Alpha and NVIDIA's mega-deals. - Gabe Goodhart defines the overarching corporate strategy behind the deals: “I think this is a clear line in the sand from NVIDIA that they are going to depend, their future is going to depend on a strong, thriving, many model ecosystem.” - Gabe illustrates Hugging Face's architectural dominance over runtime projects: “It is the reference architecture of literally every LLM on the planet that is released in open source.” - The Microsoft-GitHub precedent is cited as the best-case model for maintaining an open platform's credibility despite closed-source parent company ownership.

---

`05:00` Closed Hardware, Open Software: The Power Dynamics of AI Consolidation

Aili McConnon pushes the panel to evaluate who truly benefits in this arrangement—NVIDIA or the broader open-source community—and what underlying hazards might emerge. Ash Minhas underscores that open-source software and open model weights in NVIDIA's world still funnel down to a singular, proprietary choke point: hardware. Even as Hugging Face serves as a multi-vendor hub featuring diverse third-party inference providers, nearly all of those third parties are merely reselling compute executed on NVIDIA silicon. NVIDIA's strategy with both Hugging Face and the Poolside model factory is simply to maximize the sheer velocity of models designed, trained, and served on CUDA cores. Gabe adds that unlike hyperscalers or neoclouds—who primarily care about metering API consumption and controlling software gateways regardless of underlying silicon—NVIDIA is indifferent to who provides the hosting, so long as every server rack is packed with its accelerators.

The discussion shifts to competitive dynamics and whether NVIDIA is hedging against frontier labs developing custom in-house silicon (such as Google's TPUs, Meta's MTIA, or custom ASICs). Gabe observes that if the largest labs vertically integrate and slow down their external GPU orders, NVIDIA needs an expansive base of commodity enterprise buyers and data centers to absorb hardware volume. Tim Hwang plays devil's advocate, questioning whether commodity chip competitors were ever realistically going to catch up anyway, suggesting that critics might just need to "get real" about NVIDIA's insurmountable lead.

Gabe acknowledges that current supply-and-demand realities heavily mask these tensions. Because AI demand vastly outstrips GPU production capacity, data centers facing nine-month waitlists for NVIDIA hardware might buy chips from competing vendors just to soak up immediate demand. However, if aggregate market demand cools, an all-out turf war will emerge. The panel also contrasts full acquisitions with alternative deal structures, such as NVIDIA's deal with Poolside, which involved securing a non-exclusive license for their model factory and hiring talent rather than executing a full buyout, illustrating NVIDIA's multifaceted playbook for asserting control over AI infrastructure.

- Ash Minhas crystallizes NVIDIA's monetization thesis: “Nvidia's play here is to get as many models trained and running on their CUDA cores.” - Gabe Goodhart explains the unique alignment separating NVIDIA from hyperscalers: “NVIDIA doesn't care. It cares about the chip under the hood.” - Gabe outlines the defensive motive against frontier labs pursuing custom silicon: “I think one of the articles I read about this clearly positioned this as a defensive play against all of the labs that are trying to do vertical integration and build their own chips because if the big labs stop needing to buy nvidia chips nvidia needs to sell their chips somewhere else and so they're really going to aim at that commodity hardware like go broad rather than deep type of sales.” - Tim Hwang challenges the panel's skepticism regarding competition: “Is there another view here which is like, get real, right? Basically, these commodity players were never, ever going to catch up.” - Gabe points out that non-NVIDIA silicon vendors are currently benefiting from supply constraints: “And I also think it's important to note that the smaller hardware vendors that are trying to do commodity hardware that are not NVIDIA, they're trying to soak up excess demand, frankly.”

`10:00` NVIDIA’s Open Consolidation Strategy: Up the Stack or Let a Thousand Flowers Bloom?

The discussion kicks off with the mechanics of NVIDIA's massive investments, specifically touching on their non-exclusive licensing arrangement with Poolside. While normal companies spending $19 to $20 billion might find themselves short on cash, NVIDIA operates on an entirely different scale. Host Tim Hwang asks the panel whether NVIDIA intends to make another acquisition play in open-source AI consolidation or take a pause. Ash Minhas outlines the classic AI stack—chips and infrastructure at the base, cloud providers in the middle, and the application layer at the top. With their infrastructure layer already locked down, NVIDIA is targeting the model layer to bolster models like Nemotron using Poolside's approach to model creation. Ash wonders aloud whether NVIDIA plans to move even higher up the stack:

> “I know I'm responding to your question, Tim, with a question, but the question is, are they going to go even higher up the stack and start looking at some of that application layer and maybe going for a slice of that somewhere?”

Ash notes that NVIDIA took early steps in this direction with NeMo Guardrails / NemoClaw, making them one of the primary backers of an open-source agent harness, though generalist frameworks like Hermes have recently taken up more of the industry's oxygen. Gabe Goodhart counters that NVIDIA is unlikely to pivot toward consumer-facing, commercial application software like an `ai.nvidia.com` intended to compete directly with ChatGPT or Claude. For Gabe, NVIDIA's ultimate bottom line remains hardware:

> “I think for me the two recent deals we're talking about here are just a clear sign that they are going for as many flowers blooming as possible right and if they can put their money and their mouth behind amplify that goal, they will.”

Gabe emphasizes that NVIDIA wants to fertilize the entire ecosystem rather than wall off a narrow consumer corner and tell users to ignore the rest of the market.

```mermaid flowchart TD accTitle: NVIDIA's Position Across the AI Stack accDescr: Three-layer stack showing NVIDIA's hardware core, open model consolidation, and minimal presence at the application layer.

subgraph AppLayer["Application Layer"] Apps["Consumer Chat / Agent Harms (Hermes, Open WebUI)"] NemoClaw["NVIDIA NeMo Harness (NemoClaw)"] end

subgraph MidLayer["Model & Cloud Layer"] Poolside["Poolside Model Creation Tech"] Nemotron["Nemotron Models"] end

subgraph InfraLayer["Infrastructure & Hardware Layer"] Chips["NVIDIA GPUs / Custom Hardware (Core Business)"] Cloud["Cloud Provider Partnerships"] end

Chips --> Nemotron Poolside --> Nemotron Nemotron -.-> NemoClaw Apps -.-> Chips ```

---

`13:14` Hot Chips Reveal: IBM and Arm’s Dual-Architecture Mainframe Processor

Tim Hwang pivots the conversation from the fast-and-loose consumer AI landscape to the mainframe world, where the standard startup ethos of "just launch it and see what happens" is completely unacceptable due to mission-critical reliability demands. Aili McConnon introduces IBM's major announcement at the Hot Chips conference: a novel dual-processor architecture engineered jointly with Arm. Where frontier AI labs spent Hot Chips showcasing proprietary custom accelerators, IBM revealed a chip that integrates IBM Z mainframe enterprise workloads alongside the Arm architecture on a single piece of silicon.

```mermaid flowchart LR accTitle: IBM Dual Architecture Concept accDescr: Diagram illustrating IBM Z architecture running alongside Arm on a single piece of silicon with nanosecond switching.

subgraph DualChip["Single Silicon Die"] direction TB IBMZ["IBM Z Core<br/>(z/OS & Mission-Critical Data)"] ArmCore["Arm Core<br/>(Modern AI & Native Linux Workloads)"] Switch["Nanosecond Architecture Switching"] IBMZ <--> Switch Switch <--> ArmCore end

Inference["Local Enterprise AI Inference"] --> ArmCore BankingData["Core Banking / Gov Records"] --> IBMZ ```

Crucially, the chip is designed to switch execution between IBM Z and Arm workloads in nanoseconds. Aili explains that modern enterprise AI faces an acute bottleneck: companies desperately need to execute AI inference directly next to where their critical data actually sits, which is inside legacy mainframe stores. At the same time, the vast majority of the modern open AI software ecosystem is built for Arm. Instead of shuttling sensitive, massive datasets across external networks to separate AI clusters, the dual-processor design brings the fast-moving Arm AI ecosystem directly onto the secure mainframe substrate.

---

`15:00` Instruction Set Philosophies: RISC vs. CISC Collide in Banking Infrastructure

Skyler Speakman breaks down the fundamental computer science divergence underpinning this hardware partnership: the decades-long clash between simpler, reduced instruction sets (RISC, represented by Arm) and complex, single-instruction executions (CISC/mainframe legacy). Under one approach, adding two numbers requires distinct instructions to read from memory, perform the addition, and write back the result; under the other, a single complex instruction executes the entire sequence end-to-end. Skyler points out that putting both paradigms on one chip represents an enormous engineering feat, but it introduces deep architectural questions for highly regulated industries:

> “On the one hand, it's cool to see these fences being brought down, but there's also this approach where you should be asking, why were those fences there to begin with?”

Skyler questions the practical operational realities of merging these environments: - Will core banking institutions essentially download updates from an "App Store" style ecosystem? - Could an ATM transaction stall because backend infrastructure is occupied training or fine-tuning an AI model for 20 seconds? - Is it genuinely prudent to allow fast-moving, rapidly changing Arm software to reside in the exact same physical runtime as mission-critical systems that have quietly run government services and financial clearings for decades?

Aili McConnon echoes that for established enterprises, ripping and replacing mainframes was never an option. Opening an "App Store of options" allows enterprises to execute inference right at the data repository without forcing an impossible migration. She views the partnership as a win-win rather than a zero-sum battle between IBM and Arm, expressing pleasant surprise at the swift turnaround from initial partnership announcements to functional silicon reveal.

---

`18:40` Bulletproof Silicon vs. COBOL: The Engineering Reality of Mainframe Reliability

Gabe Goodhart steps in with the perspective of a software engineer, arguing that while industry commentators periodically predict the death of the mainframe, the underlying hardware necessity has never changed. While critics often fixate on the difficulty of maintaining legacy COBOL codebases, COBOL was merely a means to an end; the true value proposition of the IBM Z series is absolute, unmatched fault tolerance.

> “Like I was talking to I forget who it was earlier this week, who was literally part of the team that said you could literally take the mainframe out back and put a bullet through it and it would not stop processing transactions.”

Gabe points out that on an IBM Z system, an engineer could physically shoot a bullet through a RAM stick and the machine would not even blink—a level of resilience that commodity server hardware cannot match. While finding engineers capable of writing and maintaining legacy COBOL continues to be an acute pain point for corporations, integrating Arm execution directly into the mainframe allows modern developers to bring contemporary software stacks straight into this bulletproof environment without compromising its foundational reliability.

`20:00` Bringing Modern Code to the Mainframe: The Arm Advantage

Gabe Goodhart explains the core philosophy behind IBM's dual-architecture approach: enterprise reliability requirements have never disappeared, but rather than extracting legacy mainframe code out into modern environments and sacrificing that hard-won stability, the better path is bringing modern code into the mainframe. He highlights the sheer agony of ecosystem-targeted cross-compilation, pointing out that even a straightforward Python command like `import torch` actually delegates underneath to underlying C libraries tied to specific instruction sets. Before this collaboration, taking a massive LLM with complex acceleration logic and porting it to an isolated, firewalled chip like an IBM Z processor could easily swallow years of specialized engineering effort. By integrating Arm—which already runs natively across millions of smartphones, laptops, desktop processors, and cloud servers—IBM essentially inherits an enormous, pre-adapted software ecosystem:

> “Now, with Arm and the fact that a huge number of people are running Arm workloads on their laptops, on their phones, even starting to be on their desktop processors and server processors, means that a huge amount of software has already been adapted to the Arm ecosystem. So that work is done for us.”

Looking out toward the horizon of 2050, Tim Hwang asks Ash Minhas whether mainframes will genuinely endure or if the narrative that they are perpetually on their way out will finally materialize. Ash pushes back, arguing that for ultra-low latency, mission-critical transactions, mainframes remain unmatched. Latency concerns that plague distributed cloud systems vanish when processing transactions in fractions of a millisecond on a mainframe. From a pure hardware architecture perspective, Ash notes that IBM and Arm did not simply drop separate Arm cores onto the package—they fully combined the architectures into one chip, prompting questions about how running demanding frameworks like PyTorch alongside millisecond-level transactional workloads will affect overall system behavior.

```mermaid flowchart TD accTitle: Mainframe Integration Pathways accDescr: Diagram contrasting traditional bespoke porting with the Arm unified architecture pathway.

subgraph Legacy["Traditional Mainframe Porting"] A["Modern Code / PyTorch"] -->|"Years of manual cross-compilation"| B["Firewalled IBM Z Architecture"] B --> C["High Risk of Sacrificing Reliability"] end

subgraph ArmCollab["Unified Dual-Architecture Approach"] D["Widespread Arm Ecosystem"] -->|"Pre-adapted software stack"| E["Combined IBM Z + Arm Processor"] E --> F["Native Arm Linux + Low-Latency z/OS"] end ```

- **Cross-Compilation Bottlenecks**: Python machine learning libraries fundamentally depend on native C runtimes, creating immense porting friction for bespoke enterprise chipsets. - **Transactional Latency**: Mainframes excel at near-instant, millisecond transaction processing, preserving their relevance well toward 2050. - **Architectural Integration**: Rather than attaching discrete coprocessor Arm cores, the design integrates the capabilities directly into the core processor fabric.

---

`24:34` Mystery Model OX Alpha and the GLM 5.3 Reveal

Tim shifts the conversation to a story straight out of a modern cyberpunk novel: the sudden appearance on OpenRouter of a mysterious model dubbed “OX Alpha.” The model stirred intense buzz across the AI community due to its sharp reasoning capabilities and an astonishing claim in its release materials that it was capable of serving 100 trillion tokens per day. The sheer scale provoked rampant speculation about the underlying compute cluster required to sustain such throughput. Just as the panel sat down to record, the veil dropped: the mystery model was revealed to be GLM 5.3, an open-source model from China.

Ash shares that roughly a year prior, he built a personal benchmark suite to evaluate model behavior and ran OX Alpha through OpenRouter while it was unmasked:

> “And I thought, you know what? This is actually pretty good. Pretty good—I mean, it's great that it's free. It's not free now, but it's still dirt cheap, and I was pretty, pretty impressed with the output that I got.”

Gabe echoes the positive impression after giving it his own “smell test,” though he jokes that he cannot claim a true connoisseur's palate because the vast majority of his day-to-day workloads are happily satisfied by sub-30-billion-parameter models running locally on his own hardware. Still, he notes that intense competition at the frontier is thrilling to watch. He remains intensely curious about whether the 100-trillion-token metric stems from brute-force compute scaling or an architectural breakthrough in how tokens navigate the attention mechanism.

- **OX Alpha Identifier**: Stealth LLM listed on OpenRouter, ultimately unveiled as GLM 5.3. - **Throughput Claim**: 100 trillion tokens served per day, prompting debate over model architecture versus pure infrastructure scale. - **Pricing Shift**: Initially offered free during the evaluation phase, later moving to a low-cost, “dirt cheap” paid tier.

---

`27:31` The Art of the Stealth Launch: Avoiding Bias and Timing Battles

Skyler Speakman turns the focus to the psychology and marketing strategy behind launching a frontier model undercover. He questions whether anyone would be talking about a standard, scheduled release of GLM 5.3 if it had rolled out without the theatrics. The anonymity created a multi-day internet spectacle akin to a masked knight entering a medieval joust, driving organic engagement and speculation across social channels.

Gabe highlights just how treacherous it has become to time an open model release. IBM had just launched its own Granite 4.2 models—which the team designed specifically for enterprise utility rather than benchmark vanity—only to see Alibaba's Qwen team drop a benchmark-shattering model on the exact local hardware profile a single day later. With frontier labs like OpenAI and Anthropic occasionally dropping rival models within an hour of each other in playground-style brawls, an anonymous launch acts as a competition-proof shield:

> “So, like you said, Skyler, I mean, kudos to the team for choosing a different route to release because if this had been just another amazing Chinese model, like what a world we live in, by the way, you know, would we even bother to talk about it? But choosing a release strategy that almost was competition-proof because nobody else was doing it.”

Ash notes that the stealth deployment also offered an objective technical advantage: it prevented geographical and origin bias from skewing developer evaluations, allowing the team to collect completely unvarnished performance data from real-world usage. While genuine feedback played an undeniable part, the panel agrees the stealth strategy was at least half pure, brilliant marketing hype. Tim closes out episode 122 by thanking Gabe, Skyler, and Ash, before signing off until next week's discussion on *Mixture of Experts*.

- **Competitive Timing**: Launching anonymously sidesteps sudden preemptive releases from rival frontier labs like Qwen, Anthropic, or OpenAI. - **Unbiased Evaluation**: Concealing the model's lineage as a Chinese open-source release ensured testers judged performance purely on output quality rather than preconceived biases. - **Marketing Precedent**: The tactic mirrors earlier community experiments, such as anonymous arena deployments like Nano Banana, providing short windows of concentrated internet intrigue.

## Notable quotes > `05:00` Nvidia's play here is to get as many models trained and running on their CUDA cores. — Ash Minhas

> `05:00` the smaller hardware vendors that are trying to do commodity hardware that are not NVIDIA, they're trying to soak up excess demand, frankly. — Gabe Goodhart

## Mentioned Tim Hwang, Aili McConnon, Gabe Goodhart, Ash Minhas, Skyler Speakman, IBM, NVIDIA, Hugging Face, Poolside, Arm, Nemotron, NeMo Guardrails, NemoClaw, Hermes, IBM Z, LinuxONE, z/OS, Hot Chips, OX Alpha, GLM 5.3, Granite 4.2, Qwen, OpenRouter, CUDA, transformers, llama.cpp, Microsoft, GitHub, GitHub Copilot, Google TPUs, Meta MTIA, ChatGPT, Claude, PyTorch, OpenAI, Anthropic

## Video description On episode 122 of Mixture of Experts, hosts Tim Hwang and Aili McConnon are joined by Gabe Goodhart, Ash Minhas, Skyler Speakman to talk hot chips, mystery models, and major money moves. We open with IBM's Hot Chips conference reveal of its dual-architecture mainframe processor built with Arm that lets IBM Z and LinuxONE systems run Arm-native Linux workloads right alongside z/OS. We break down what this means for enterprises balancing legacy reliability with the fast-moving Arm software...

[Source](https://cdnapisec.kaltura.com/p/1773841/sp/177384100/playManifest/entryId/1_rqeaunh4/format/download/protocol/https/flavorParamIds/0/name/media.mp3)

中文简报

英伟达收购Hugging Face，将开源AI锁定于其专有GPU 在前沿实验室打造自研芯片之际，英伟达正在资助开放权重模型生态系统，以确保开发者持续依赖其专有的CUDA芯片。 讨论首先聚焦英伟达据报道对Hugging Face和Poolside进行的巨额收购与投资，分析这家硬件巨头如何利用开源软件和开放模型权重来拉动对其专有CUDA基础设施的需求，同时防范自研定制芯片的冲击。随后，讨论转向IBM在Hot Chips大会上宣布的将IBM Z与Arm整合在同一芯片裸片上的双架构芯片，探讨融合CISC与RISC执行能力如何让关键任务型企业在不牺牲坚如磐石的可靠性的前提下，直接紧邻传统数据存储运行现代AI工作负载。最后，主持人深入剖析了Ox Alpha在OpenRouter上的隐秘上线（在录制前刚刚被揭露为中国的GLM 5.3），并探讨隐匿发布如何既能提供无偏见的技术测试，又能防范来自前沿AI实验室激烈的发布时间战。 英伟达正通过据报道的巨额交易迅速整合开源AI基础设施，包括以129亿美元收购Hugging Face，以及与代码生成初创公司Poolside达成的一笔60亿美元交易。 (0:00) IBM在Hot Chips大会上公布了与Arm合作开发的一款双架构大型机处理器，在单芯片裸片上集成了IBM Z的企业级可靠性与原生Arm执行能力。 (13:14) 在OpenRouter上声称日吞吐量达100万亿token的隐秘推理模型“OX Alpha”，被揭晓实为中国开源模型GLM 5.3。 (24:34)

Reading summary

On episode 122 of IBM's Mixture of Experts, panelists analyze NVIDIA's aggressive investments in open-source AI infrastructure to secure hardware demand, alongside IBM and Arm's new dual-architecture mainframe silicon and the stealth rollout of GLM 5.3. NVIDIA's Open-Source Play to Lock In Hardware Demand (0:00) NVIDIA's reported $12.9 billion acquisition of Hugging Face and a $6 billion transaction with code-generation startup Poolside reflect a corporate strategy designed to secure long-term hardware consumption. Gabe Goodhart emphasizes that NVIDIA relies on a decentralized, multi-model ecosystem rather than a single monolithic provider. Hugging Face occupies a pivotal role as the upstream reference architecture for open models via its transformers library, while Poolside provides model-creation technology to power initiatives like NVIDIA's Nemotron family. Ash Minhas points out that this open ecosystem ultimately funnels into a proprietary choke point: maximizing the velocity of models trained and served on NVIDIA CUDA cores. Tensions in Model Consolidation and Stack Positioning (5:00) While backing the open ecosystem provides financial stability, Goodhart highlights an inherent tension between the open-source community's goal of running smaller, cheaper models on heterogeneous commodity hardware and NVIDIA's incentive to sell proprietary accelerators. Furthermore, panel discussions explore whether these investments act as a defensive hedge against frontier labs vertically integrating with custom silicon like Google TPUs or Meta's MTIA. Regarding vertical expansion, Minhas notes NVIDIA has explored open agent harnesses like NeMo Guardrails, but Goodhart argues the company is unlikely to compete directly at the consumer application layer with an offering like ChatGPT, preferring instead to stimulate broad market hardware adoption. IBM and Arm Merge Enterprise Mainframes with RISC Silicon (13:14) At the Hot Chips conference, IBM revealed a dual-processor architecture co-developed with Arm that combines IBM Z mainframe workloads with Arm cores on a single piece of silicon, switching execution within nanoseconds. This design addresses the operational need to run modern enterprise AI inference directly beside sensitive legacy data repositories without moving petabytes over external networks. Skyler Speakman notes the computer science contrast between Arm's reduced instruction set (RISC) and mainframe complex instruction sets (CISC), raising operational questions about whether co-locating fast-moving modern software on mission-critical banking infrastructure risks system stability. Aili McConnon and Goodhart counter that mainframes remain irreplaceable for low-latency fault tolerance, arguing that bringing Arm's vast, pre-adapted software ecosystem directly to the mainframe avoids years of bespoke cross-compilation. The Strategic Dynamics of Anonymous Model Releases (24:34) The panel reviewed the sudden emergence of 'OX Alpha' on OpenRouter—a stealth reasoning model claiming throughput of 100 trillion tokens per day—which was subsequently revealed to be China's open-source GLM 5.3. Minhas and Goodhart report favorable performance on evaluation suites, noting its initial free availability before moving to a low-cost tier. Speakman and Goodhart argue the anonymous deployment was both an effective marketing maneuver and a tactical shield against sudden counter-releases by frontier labs such as OpenAI, Anthropic, or Alibaba's Qwen. In addition, Minhas notes that stripping brand identity allowed the developers to gather objective user evaluation data unmarred by geographical or origin bias. The discussions leave open whether NVIDIA's dominance will face pressure once current supply constraints ease and frontier labs scale in-house silicon, as well as how smoothly enterprises can integrate modern AI frameworks alongside mission-critical mainframe workloads without introducing transactional latency or reliability risks.

精炼阅读摘要

本期节目探讨了企业在推动开源生态繁荣与底层专有硬件绑定之间的权衡，并展示了通过跨架构融合和隐蔽发布策略在激烈AI竞争中突围的行业新路径。 NVIDIA推进开源收购以巩固硬件底座 (0:00) NVIDIA近期开展了密集的资本布局，包括以129亿美元收购Hugging Face的传闻，以及与AI代码生成初创公司Poolside达成的一笔60亿美元交易（包括技术非独占许可及人才引入）。针对这些动作，嘉宾Gabe Goodhart指出，NVIDIA的核心商业模式并不依赖单一的封闭模型供应商，而是依赖庞大且去中心化的开源AI生态。Hugging Face不仅是模型托管平台，其Python库`transformers`更已成为几乎所有开源大模型的实际参考架构。例如，下游项目（如`llama.cpp`）通常会搁置或拒绝架构特性的合并请求，直到该特性率先并入Hugging Face的上游仓库。 然而，这一策略引发了激励机制上的潜在冲突。开源社区长期致力于将模型缩小、降低成本，使其能够在异构边缘硬件和通用芯片上运行；而NVIDIA的核心商业利益则是销售昂贵的专有算力。对此，Ash Minhas与Gabe分析认为，无论模型如何开源或多供应商托管，当前的推理与训练流量最终仍需流向底层的CUDA核心。NVIDIA此举旨在加速模型开发与运行的速度，同时也是一种防御性策略，以防前沿实验室（如自研TPU的Google或MTIA的Meta）实现垂直整合并缩减GPU订单时，NVIDIA能通过更广泛的通用企业买家和数据中心来消化产能。 IBM与Arm打造双架构大型机芯片 在Hot Chips大会上，IBM联合Arm发布了一款新型双处理器架构芯片，首次在单块硅片上集成了IBM Z大型机企业工作负载与Arm架构，并宣称能在纳秒级别内完成任务切换。Aili McConnon解释道，现代企业AI面临的核心瓶颈在于：企业极度需要在存放关键业务数据的传统大型机内部就近执行AI推理，但现存的大多数现代开源AI软件生态（如各类针对Arm高度适配的工具链）与封闭的大型机指令集难以直接兼容。该设计旨在将快速演进的Arm软件栈直接引入安全、高可靠的大型机环境中，免去通过外部网络搬运敏感海量数据的风险。 从计算机科学角度来看，Skyler Speakman指出了RISC（精简指令集，以Arm为代表）与CISC（复杂指令集，代表传统大型机）在同一芯片上的碰撞。尽管拆除架构隔离展示了极高的工程造诣，但也带来了监管与系统调度层面的疑问：例如银行在同一硬件底座上运行20秒的AI微调任务，是否会干扰对延迟极度敏感的核心ATM或跨行清算交易。不过，Aili和Gabe指出，对于金融和政府机构而言，彻底淘汰并重构大型机既不现实也不可行，大型机具备如物理损坏依然保证高容错的硬件冗余，通过无缝引入Arm生态，使现代开发者无需耗费数年重写C底层的交叉编译逻辑，便能在坚固的底座上直接运行现代模型。 OX Alpha隐蔽发布的策略与竞争考量 (24:34) 此前在平台OpenRouter上代号为“OX Alpha”的神秘模型引发了行业热议。该模型凭借出色的推理能力以及宣称每天可处理100万亿Token的惊人吞吐量，引发了对底层集群规模和注意力机制架构创新的广泛猜测；直到录制前夕，该模型才被揭晓为中国的开源模型GLM 5.3。Ash分享了自己的基准测试体验，指出该模型在最初免费试用及后续极低收费阶段的输出表现均令人印象深刻。Gabe也对其表现给予了正面评价，但同时提出疑问：每日100万亿Token的庞大指标究竟是通过蛮力算力集群实现，还是依赖某种特殊的架构优化，目前尚无定论。 针对这种“匿名空降”的策略，Skyler与Gabe探讨了其背后的竞争心理学与营销技巧。在开源大模型频出的当下，各家实验室的发版节奏高度重叠且竞争激烈（例如IBM发布专注企业级实用的Granite 4.2后，阿里Qwen团队次日便发布了同规格刷榜模型，甚至OpenAI与Anthropic经常在数小时内互相抢发）。隐蔽发布不仅能制造出如同中世纪骑士蒙面比武般的社交话题度，规避同期的发版冲突，Ash更强调，隐藏模型背景与地域信息能有效消除开发者的地域和来源偏见，从而在盲测环境下收集到真实、无偏见的用户反馈数据。 本期讨论表明，AI基础设施的竞争正向底层硬件与生态控制权双向延伸：NVIDIA试图通过开源入口锁定CUDA硬件需求，IBM尝试通过架构融合将现代AI引入高容错大型机，而新兴模型则开始采用隐蔽测试绕过发版踩踏与认知偏见。至于异构架构能否在高可靠行业完全化解稳定性风险，以及非NVIDIA芯片能否在供需缓解后打破垄断，仍有待行业长期验证。

Detailed notes — zh

## Background 该音频由 IBM 于 2026 年 8 月 28 日在其播客栏目《Mixture of Experts》（第 122 期）发布。根据页面介绍，主持人 Tim Hwang、Aili McConnon 携手嘉宾 Gabe Goodhart、Ash Minhas 与 Skyler Speakman，深入探讨了 IBM 在 Hot Chips 大会上曝光的双架构大型机芯片、英伟达重金布局 Hugging Face 和 Poolside 背后的开源野心，以及引发广泛猜测的神秘模型 Ox Alpha 的幕后真相。 ## TL;DR 本期节目聚焦 AI 领域最新的资本运作与底层硬件颠覆性变革。嘉宾们首先剖析了英伟达豪掷百亿美元收购 Hugging Face 与授权 Poolside 的深层逻辑，指出其本质是通过繁荣“开放权重与软件”生态来最终变现其封闭的 CUDA 硬件帝国。随后，对话深入探讨了 IBM 与 Arm 联合研发的双架构大型机芯片，解析其如何通过硬件级深度融合，在坚若磐石的关键交易系统上直接原生运行现代 Arm 与 AI 负载，进而终结 COBOL 的长期痛点。最后，节目复盘了在 OpenRouter 上以代号“OX Alpha”神秘亮相、最终确认为 GLM-5.3 的开源模型，探讨了其宣称的超高吞吐量以及利用匿名“盲测”既消除地域偏见又实现顶级营销的巧妙发版博弈。 ## Atmosphere & style 整期对话专业严谨却充满极客幽默，宛如一场顶尖系统工程师与行业观察者围坐展开的高能技术拆解会。主持团队与嘉宾之间配合默契，既有底层微架构与指令集层面的技术交锋，也不乏对大厂资本阳谋与宣发博弈的犀利反讽。现场气氛既有探讨关键系统“防弹级稳定性”时的敬畏与审慎，又充满了解密神秘大模型“比武招亲”般的轻快与生动，给听众带来扎实硬核又极具沉浸感的科技洞察体验。 ## Bulletin - `00:00` NVIDIA 推进以“开放”为核心的资本布局，涉及拟以 129 亿美元收购 Hugging Face 及此前向 Poolside 达成的 60 亿美元非独占授权交易，旨在通过繁荣开源软件生态间接绑定底层 CUDA 硬件的持续消耗。 - `13:30` IBM 在 Hot Chips 大会上公布与 Arm 联合研发的双架构大型机芯片，支持在单个芯片上纳秒级切换 IBM Z 原生关键任务与现代 Arm Linux/AI 负载，避免在异构系统间搬迁核心业务数据。 - `25:00` 此前在 OpenRouter 上以代号“OX Alpha”匿名测试并宣称具备每日 100 万亿 Token 吞吐能力的神秘模型正式揭晓，实为中国开源大模型 GLM-5.3。 - `05:00` 当前芯片市场供需严重失衡，部分数据中心采购 NVIDIA 芯片面临长达 9 个月的提货排期，这为其他硬件供应商提供了短期吸收外溢算力需求的空间。 ## Key takeaways - `05:00` NVIDIA 围绕开源生态大举注资的核心战略是“开放权重、开放软件，但封闭硬件”，无论模型与分发网络如何多样化，终极变现途径均是物理芯片与算力卡。 - `00:00` Hugging Face 的 Transformers 软件库已成为开源 AI 领域的事实架构标准，在 Llama.cpp 等下游开源项目中，未经上游合并的代码往往直接被拒绝采纳。 - `00:00` 开源生态追求在廉价通用硬件上运行精简模型的初衷，与芯片厂商追求销售昂贵大算力硬件的诉求存在内生激励冲突。 - `05:00` 面对头部前沿 AI 实验室自研芯片的垂直整合趋势，NVIDIA 通过大举投资开源模型工厂和开发者生态，采取全行业“广撒网”的防御性护城河策略。 - `10:00` 分析指出 NVIDIA 极不可能亲自下场做对标 ChatGPT 或 Claude 的 C 端通用对话应用，其技术栈渗透主要停留在中间层与模型工厂，以维持广泛客户基础。 - `13:30` 企业级 AI 落地的核心诉求是将推理算力尽可能移至核心业务数据存放处，而全球海量关键事务数据依然沉淀在 IBM 大型机环境中。 - `15:00` IBM 与 Arm 的双处理器融合实现了复杂指令集（CISC）与精简指令集（RISC）在微架构层面的硬件级融合，打破了长久以来的系统物理隔离。 - `20:00` 将 Arm 架构引入大型机芯片改变了以往强行将关键业务代码外迁的模式，直接利用已适配手机、PC 及服务器的成熟 Arm 软件生态避免耗时数年的交叉编译。 - `18:00` 大型机所具备的极高容错设计与毫秒级交易处理能力在现代计算中仍具不可替代性，未来核心系统更可能演进为一体化混合超级专用机器，而非彻底消亡。 - `25:00` “OX Alpha”（GLM-5.3）所宣称的每日 100 万亿 Token 极高吞吐量引发了行业关注，核心悬念在于其依赖纯算力堆叠还是在注意力机制等架构层面实现了算法创新。 - `27:30` 开源模型采取匿名“盲测”上线策略既能剔除地域与品牌带来的认知偏见、获取客观性能反馈，也是在激烈的宣发窗口博弈中避免与欧美主流大厂直接撞车的营销奇谋。 ## Detailed walkthrough ### `00:00` NVIDIA 的资本重拳：重金押注 Hugging Face 与开源生态 节目伊始，主持人 Tim Hwang 与联合主持人 Aili McConnon 欢迎本期嘉宾 Gabe Goodhart、Ash Minhas 以及 Skyler Speakman。本周 AI 领域的新闻密集到令人应接不暇，除了 IBM 在 Hot Chips 大会上公布的下一代双架构处理器、神秘模型 Ox Alpha 背后团队的曝光，最震撼业界的核心话题莫过于 NVIDIA 近期展现的野心勃勃的收购与投资动作——包括以 129 亿美元收购 Hugging Face 的重磅消息，以及此前与 Poolside 达成的 60 亿美元交易。 AI Foundations 首席架构师 Gabe Goodhart 敏锐地指出，这两笔交易的核心逻辑完全围绕着“开放”展开。NVIDIA 正在用真金白银表明立场：其硬件帝国未来的繁荣，完全寄托在一个生机勃勃、百花齐放的多模型生态系统之上，而当前这个生态唯有通过开源来实现。从推进自家的 Nemotron 模型系列，到借由 Poolside 的投资打造可靠的模型工厂，再到直接拿下 Hugging Face，NVIDIA 正在一步步走向极致。Hugging Face 如今不仅是开源 AI 的基石，其旗下的 Transformers 软件库更是整个行业的事实标准和参考架构。Gabe 举例称，在日常工程实践中，如果你试图向 Llama.cpp 提交一个未经 Hugging Face Transformers 仓库合并的 PR，维护者甚至会直接拒绝合并。NVIDIA 此举显然是在争夺开源 AI 的“重力中心”。 这种巨额资本的注入也带来了一把双刃剑。有 NVIDIA 这样资金雄厚、地位稳固的巨头背书，开源社区未来数年都将拥有强大的财力支持；但另一方面，开源生态的一大驱动力是让模型更小、更通用，能够跑在各种廉价硬件上，而 NVIDIA 作为商业芯片厂商，其核心诉求是推销昂贵的大算力硬件，这不可避免地滋生了某种扭曲的激励机制。业界普遍将其与微软当年收购 GitHub 进行对比，普遍期望 NVIDIA 能够像微软守护 GitHub 的开源信条一样，为 Hugging Face 保留纯粹的开源底色。Gabe Goodhart 总结道：“This is a clear line in the sand from NVIDIA that they are going to depend, their future is going to depend on a strong, thriving, many model ecosystem.”（这是 NVIDIA 划下的一道明确底线，表明他们的未来将依赖于一个强大、繁荣的多模型生态系统。） ```mermaid flowchart TD accTitle: NVIDIA 开源生态控制链 accDescr: 展示 NVIDIA 如何通过软件标准和模型层最终导向其底层封闭硬件销售 subgraph SW ["软件与开源层"] HF["Hugging Face (Transformers 架构标准)"] Poolside["Poolside 模型工厂 & Nemotron 系列"] end subgraph Infra ["基础设施与分发"] Neoclouds["各大推理服务商 / NeoClouds"] OpenCommunity["开源社区与企业应用"] end subgraph HW ["底层硬件变现"] CUDA["NVIDIA 芯片 (CUDA 核心销售)"] end HF --> OpenCommunity Poolside --> HF OpenCommunity --> Neoclouds Neoclouds -->|"采购与部署"| CUDA ``` ### `05:00` 开放权重、封闭芯片：谁才是真正的赢家？ Aili McConnon 随后抛出了一个核心疑问：在这场豪赌中，最大的受益者究竟是开源社区还是 NVIDIA 自己？Skyler Speakman 和 Ash Minhas 随即加入讨论，直击这笔交易的商业本质——“开放权重、开放软件，但封闭硬件”。在他们看来，无论上层应用如何繁荣，NVIDIA 唯一的终极目标就是让尽可能多的模型在自己的 CUDA 核心上训练和运行。只要终端运行依赖的是 NVIDIA 的硬件，它甚至毫不介意 Hugging Face 接入各家第三方推理服务商，因为那些云服务商最终同样要向 NVIDIA 采购芯片。Skyler Speakman 尖锐地指出：“Open weights, open software, but closed hardware.”（开放权重，开放软件，但硬件是封闭的。） 这一策略同样也是针对前沿 AI 实验室自研芯片潮流的一记防御性反击。随着头部大模型实验室纷纷走向垂直整合、试图摆脱对 GPU 的依赖，NVIDIA 必须防止自身客户面变窄，转而走“广撒网”路线，牢牢占据通用数据中心和大宗市场。在这场重新洗牌中，最大的输家将是那些试图制造大宗商品化芯片的非 NVIDIA 硬件竞争对手。面对现场关于 NVIDIA 垄断野心的担忧，Tim Hwang 扮演起了“反派角色”，直白地质疑道：这是否只是一种残酷的现实主义？毕竟那些通用芯片厂商本来就落后太多、根本无力追赶，NVIDIA 只是在顺理成章地兑现既有优势而已。 对此，Gabe Goodhart 分析认为，当前整个芯片生态仍处于极度不平衡的供需关系中，市场缺口大到数据中心的提货排期动辄长达九个月，只要买不到 NVIDIA 芯片，买家自然会寻找替代方案，因此其他硬件供应商短期内只是在吸收过剩需求。然而，一旦未来整体算力需求出现降温，更惨烈的肉搏战才会真正打响。此外，Aili 提到的 Poolside 交易模式也揭示了一种新的 AI 整合范式——NVIDIA 并未采取全盘收购，而是通过获取非独占许可（non-exclusive license）并吸纳核心团队的方式切入，以灵活的商业手段持续扩张其生态版图。 - **核心收购与交易规模**：NVIDIA 拟以 129 亿美元收购 Hugging Face，此前完成了与 Poolside 涉及 60 亿美元的相关交易。 - **开源生态的参考坐标**：Hugging Face 的 Transformers 是开源 LLM 的事实架构标准，上游合并状态直接影响如 Llama.cpp 等下游开源项目。 - **商业模式本质差异**：超大规模云厂商（Hyperscalers）看重 API 接口与计量网关（metering gateway），而 NVIDIA 只在乎底层驱动模型运行的物理芯片。 - **市场供需现状**：GPU 依然处于极端供不应求状态，部分数据中心采购 NVIDIA 芯片面临长达 9 个月的提货等待期。 ### `10:00` 英伟达的生态野心与应用层走向 在讨论英伟达与 Poolside 的授权协议时，嘉宾们分析了这笔交易的具体细节。该协议并非排他性的，这意味着 Poolside 还可以将其模型构建方法授权给其他工具类公司，但业内普遍怀疑是否有其他客户愿意掏出高达 60 亿美元的巨款来购买同样的许可。主持人 Tim Hwang 提出疑问：在接连砸下 190 亿至 200 亿美元的巨额资金后，像英伟达这样弹药极其充足的巨头，下一步还会去开源领域整合谁？他们会继续出手，还是暂时按兵不动？ Ash Minhas 随即拆解了当前的技术堆栈：底层是芯片与算力基础设施，中间是云服务提供商与基础模型，最上层则是具体应用。英伟达此前已经牢牢锁定了基础设施层，现在正通过 Poolside 等交易强势渗透中间层，以改良其自家的 Nemotron 模型。对此，Ash 提出了一个反问：“问题在于，他们会不会进一步向技术栈上层走，开始盯上部分应用层，并尝试在某个领域分一杯羹？”他们是否会直接去收购 Open WebUI 这样的产品，甚至以 NemoClaw 这类开源智能体（Agent）套件为跳板更进一步？ Gabe Goodhart 对此持怀疑态度。他指出虽然英伟达凭借 NemoClaw 成为开源智能体底座的重要推手，但近期围绕它的热度有所降温，像 Hermes 这样的通用智能体框架抢占了更多关注度。Gabe 强调，英伟达绝不可能去走直接面向 C 端的商业软件路线，大家不会看到一个类似 `ai.nvidia.com` 的平台让用户像使用 ChatGPT 或 Claude 那样去聊天。正如 Gabe Goodhart 所言：“这在我看来，他们最近讨论的两笔交易明确表明，他们希望看到‘百花齐放’；如果他们能出资发声来放大这个目标，他们就会这么做。”英伟达的核心商业逻辑始终紧紧绑定在底层芯片的消耗上，他们绝不会去划地为牢、逼着用户只用自家的单一应用。 - 交易特性：Poolside 与英伟达的协议为非独占授权，但 60 亿美元的高门槛让后续跟进者寥寥无几。 - 技术栈布局：英伟达在完全巩固基础设施后，正通过开源模型（如 Nemotron）与中间层工具强化中端生态。 - 智能体生态：开源 Agent 框架竞争激烈，NemoClaw 声量近期受到 Hermes 等通用框架的挤压。 ### `13:30` Hot Chips 大会重磅：IBM 与 Arm 联手打造双架构大型机芯片 随后，Tim 将话题转向了逻辑完全不同的大型机（Mainframe）市场。与主流 AI 行业“先发布上线、看看会发生什么”的激进试错文化截然相反，大型机领域容不得半点差错。Aili McConnon 介绍了本周在 Hot Chips 大会上曝光的重磅消息：在前沿芯片厂商竞相发布自研芯片的背景下，IBM 联合 Arm 推出了一款颠覆性的双处理器架构芯片。该芯片让 IBM Z 体系的大型机原生负载与 Arm 架构软件得以在同一块芯片上并存，并且系统能在纳秒（nanosecond）级别实现两种工作负载的极速切换。 这一设计的背后，直指当前企业级 AI 落地最核心的痛点：企业迫切希望将 AI 推理尽可能靠近核心业务数据存放的地方，而全球大量关键数据正是沉淀在 IBM 大型机系统中；与此相对的是，现代主流 AI 软件生态有很大一部分是建立在 Arm 架构之上的。IBM 的策略并非在异构系统间大费周章地搬运庞大且敏感的数据，而是反其道而行之，通过物理硬件的融合，将前沿软件栈直接“搬”到核心数据跟前。 ```mermaid flowchart TD accTitle: IBM Z 与 Arm 双架构协同示意 accDescr: 展示企业数据在同一芯片上实现大型机关键任务与现代 Arm AI 推理的纳秒级无缝切换 subgraph DualChip["IBM 双架构处理器 (Single Chip)"] direction LR ZCore["IBM Z 架构<br/>(运行 z/OS / 核心事务)"] ArmCore["Arm 架构<br/>(运行现代 AI 推理 / Linux)"] Switch["纳秒级工作负载切换机制"] ZCore <--> Switch Switch <--> ArmCore end Data[("企业核心任务数据<br/>(银行、政务、金融)")] --> DualChip DualChip --> Output["原地完成高安全业务与低延迟 AI 推理"] ``` ### `15:00` CISC 与 RISC 的碰撞：拆除安全护栏的隐忧与机遇 针对这一架构创新的深层影响，Skyler Speakman 从计算机指令集哲学的底层演进展开了剖析。历史上计算机科学分化出两种截然不同的流派：一种是复杂指令集（CISC），用单条复杂指令端到端完成从内存读取、相加到写回的所有操作；另一种则是精简指令集（RISC，如 Arm），倾向于拆解成多条简单独立的指令协同执行。如今这两大流派在同一块芯片上融为一体，硬件底层实现的精妙程度令人惊叹。 但 Skyler 也提出了尖锐的质疑与现实担忧：这种跨界融合会不会对关键系统的稳定性构成冲击？如果未来银行也像应用商店一样频繁更新核心软件，或者用户在 ATM 机取款时被提示“后端正在训练模型，请稍等 20 秒”，这将彻底打破大型机以往的确定性。正如 Skyler Speakman 所指出的：“从一方面看，看到这些藩篱被打破确实很酷，但你也必须反思，那些藩篱最初为什么会被立在那里？我们真的希望让 Arm 软件与数十年来看护政府和银行服务的核心系统坐在一起吗？” Aili 对此回应表示，对于无法承受彻底推倒重来的大型企业而言，这种设计就像是给大型机用户直接接入了一个庞大的现代应用商店，让他们不必搬移数据就能获得现代化推理能力。这是一场 IBM 与 Arm 的双赢合作，而非零和博弈。 - 指令集哲学碰撞：CISC（端到端复杂指令）与 RISC（精简多指令协同）在微架构层面实现硬件级合流。 - 架构切换性能：IBM Z 原生指令与 Arm 指令集支持纳秒级切换，消除了跨网络或跨设备传输数据的瓶颈。 - 潜在系统风险：关键任务（Mission-Critical）运行环境面临灵活性与极端安全稳定之间的严峻权衡。 ### `18:00` 软件工程视角：终结 COBOL 痛点与无坚不摧的硬件韧性 从软件工程的实操视角出发，Gabe Goodhart 指出这种双架构融合将极大解放软件工程师的生产力。过去市场曾一度恐慌大型机是否会随着 COBOL 语言的人才凋零而走向消亡，但 IBM 早就给出了极其清晰的回应：COBOL 语言只是一种实现手段，而企业对极致高可用性与可靠性的终极需求从未改变。 为了阐述大型机近乎神话般的容灾能力，Gabe 分享了他此前与大型机团队成员的交流轶事：在 IBM Z 系统上，其容错设计甚至达到了“你把大型机搬到后院开枪射击，哪怕直接打穿一根内存条（RAM stick），它都不会眨一下眼，核心交易处理依然不会中断”的离谱程度。目前市面上没有任何其他商用硬件能够企及这种级别的物理容错率。 Gabe 总结道，全球诸多最关键的基础设施依然深深扎根在这种坚如磐石的硬件基座上。尽管海量核心逻辑仍被困在古老且维护成本极高的 COBOL 代码中、懂这门语言的程序员也日渐稀缺，但 Arm 生态的原生接入，恰恰让现代开发者能够在不牺牲大型机底层“防弹级”稳定性的前提下，使用前沿现代工具栈重新焕发这些核心系统的生命力。 ### `20:00` 大型机融合 Arm 架构的底层逻辑与 2050 年的技术走向 Gabe Goodhart 深入剖析了企业对大型机极致可靠性（reliability）的刚性需求从未消退这一现实。过去的传统做法要么是尝试把大型机代码迁出，但这不可避免地会牺牲企业赖以生存的高可用性与稳定性；而 IBM 此次与 Arm 合作的核心思路则是“反其道而行之”——与其把代码迁出，不如将现代代码直接引入大型机内部。Gabe 分享了自己为各种冷门或专用架构交叉编译代码的痛苦经历：表面上看像 `import torch` 这样简单的 Python 程序，其底层本质上全都是 C 语言编写的底层实现，任何库都在向下委托给特定架构的 C 库，生态目标下的交叉编译堪称巨大的噩梦。在这次发布之前，把大语言模型（LLM）这种背后充满复杂加速逻辑的代码完整交叉编译并移植到与外界隔离的专用 Z 系统芯片上，往往需要数年时间。然而，如今大量开发者已经在笔记本、手机、桌面及服务器处理器上广泛运行 Arm 负载，海量软件已经天然适配了 Arm 生态。将 Arm 引入大型机芯片意味着繁重的基础移植工作已经由庞大的社区提前完成，从而打开了将海量前沿工作负载轻松引入大型机的大门。不过，Gabe 和 Skyler Speakman 也指出，多架构共享处理器后究竟会对原核心工作负载的高可靠性产生何种影响，依然“事实胜于雄辩”（the proof will be in the pudding）。 ```mermaid flowchart TD accTitle: 大型机引入 Arm 架构工作流 accDescr: 对比传统迁移代码的困境与引入 Arm 架构后直接运行现代负载的流程 subgraph Past["以往方案：将大型机代码外迁"] A["传统大型机 (z/OS)"] -->|"尝试将代码迁出"| B["牺牲极高可靠性"] C["现代复杂应用 (如 PyTorch)"] -->|"跨架构交叉编译 (耗时数年)"| A end subgraph Present["当下方案：将 Arm 生态引入大型机"] D["庞大 Arm 软件生态 (手机/PC/服务器)"] -->|"原生代码直接跑在 Arm 核心"| E["统一混合芯片 (IBM Z + Arm)"] E -->|"保留毫秒级低延迟"| F["兼顾极致可靠性与现代 AI 负载"] end ``` 谈及更长远的未来，主持人 Tim Hwang 抛出了一个设问：到了 2050 年，我们是否还会使用大型机？大众常常抱怨大型机技术陈旧，甚至不断唱衰它行将就木，但其无可替代的可靠性却始终立于不败之地。Ash Minhas 从底层系统与计算架构的角度给出了肯定答案：大型机在交易处理速度上拥有的绝对低延迟优势在其他架构中极难复现，处理毫秒级核心交易的需求将永远存在，因此未来只会走向更大、更强大的专用机器。Ash 进一步提出了架构层面的困惑：IBM 在芯片实现上并未简单地外挂几个 Arm 核心，而是采用了深度一体化融合的方式。Ash 对这种艰难的工程选型充满了好奇，并直言：“很想找做出这一架构决策的人聊聊，问他们为什么要选择走这条更艰难的路，因为换一种方式本来可以轻松得多。”同时，如果在处理核心交易的同时并发运行 PyTorch 等繁重负载，系统整体表现会受到何种影响，这也将决定未来的企业是否会在 2050 年需要一台规模更为庞大的巨型计算机。 - 企业核心诉求：与其冒着失去容灾与稳定性的风险将传统关键业务移出大型机，不如将成熟的外部开源生态直接接纳进来。 - 软件工程壁垒：Python 代码底层高度依赖 C 库与系统架构，传统移植工作量往往以“年”为单位计算。 - 生态红利互通：终端设备对 Arm 的普及使得主流 AI 工具链天然适配该架构，大幅缩减了工程适配开销。 - 架构深度整合：IBM 并没有采用简单的多芯片拼接或外挂 Arm 核心方案，而是完成了深度片上架构融合。 --- ### `25:00` 迷雾中的 OX Alpha：GLM-5.3 曝光与顶级开源模型的硬实力 Tim Hwang 将话题转向了当周极具“赛博朋克”小说色彩的神秘事件：OpenRouter 平台上突然上线了一个代号为“OX Alpha”的神秘模型。该模型展示出了极其惊人的推理与生成能力，最引人瞩目的是其官方声称具备每日处理 100 万亿（100 trillion）Token 的惊人吞吐吞吐量，这立即引发了整个 AI 社区对其背后究竟调用了多大算力集群的广泛猜想。就在节目录制当天早晨，迷雾揭晓，该模型正式确认为来自中国的开源大模型 GLM-5.3。关于这款模型的综合表现，Ash Minhas 表示他此前就搭建了一套自己的基准测试集，并在 OpenRouter 上进行了实测，结果令人印象极为深刻，且调用成本极其低廉。 Gabe Goodhart 随后进行了幽默的点评，他形容品鉴模型就像品酒，而自己现在已经算不上拥有挑剔舌头的“品酒鉴赏家”了，因为他平时大部分需求仅靠本地 DGX 工作站上运行的一个 30B（300 亿参数）以下的小模型就能满足。不过，Gabe 强调顶尖梯队能够出现高水平竞争令人振奋，同时他也对 100 万亿 Token 吞吐的底层机理提出了核心技术疑问：“我非常好奇他们所声称的这一吞吐量数字，究竟是单纯靠海量算力与并行扩展的蛮力实现的，还是他们在模型架构上做出了真正的创新，从而在保持顶级模型质量的同时实现了吞吐量的几何级飞跃。”如果开发团队真的找到了让数据更快穿过注意力机制（attention mechanism）的底层算法技巧，那将是极具突破性的成就。 - 吞吐量指标：OX Alpha / GLM-5.3 对外宣称每日可承载高达 100 万亿 Token 的生成服务。 - 成本优势：模型在 OpenRouter 早期测试阶段几乎免费，后续转为商业调用后价格依旧极低。 - 边缘与本地化：Gabe 指出对于日常很多工作而言，小于 30B 参数的模型在本地 DGX 硬件上已足够胜任。 - 架构疑问：核心悬念在于极高的并发吞吐究竟是算力堆叠的结果，还是优化了 Attention 逻辑的架构创新。 --- ### `27:30` 匿名“盲测”发版策略：营销炒作还是规避偏见？ 对于为什么模型开发者选择以“OX Alpha”这种完全匿名的模式悄然上线，嘉宾们展开了热烈探讨。Skyler Speakman 认为这种隐蔽模式（stealth mode）取得了极佳的市场营销效果。试想如果只是平铺直叙地发布 GLM-5.3，或许根本无法在国际社区引发如此广泛的声量；而披上神秘面纱后，整个过程就像是一场中世纪的比武招亲，神秘骑士头戴全封闭头盔入场大杀四方，瞬间引爆了社交媒体长达数天的持续讨论与推理猜想。 Gabe Goodhart 则从模型团队的发布博弈角度进行了补充。他结合 IBM 内部的发布经历苦笑着透露，模型发布的时机把控极其残酷。IBM 刚刚发布了团队引以为傲的 Granite 4.2 模型，虽然不追求在基准跑分上一骑绝尘，但在特定目标领域极为好用；然而就在发布次日，通义千问（Qwen）紧接着就丢出了一款在相同本地硬件规格下跑分逆天的新模型，完全打乱了节奏。Gabe 指出：“挑选在何时以及如何发布模型，本身就是一场高难度的投机博弈，你得去赌还有谁会在同一个空间、同一尺寸级别发布竞争模型。”业界此前甚至出现过 OpenAI 和 Anthropic 在相差短短一个小时内先后扔出重磅模型、在聚光灯下争抢眼球的“游乐场式对轰”。因此，像 OX Alpha 这样用抓人眼球的名字进行匿名暗度陈仓的发布，成了规避与前沿大厂正面撞车的一种极具智慧的策略。 ```mermaid flowchart TD accTitle: 匿名发布策略的双重驱动力 accDescr: 展现隐藏身份上线模型在消除偏见和市场博弈上的双重考量 A["匿名上线 'OX Alpha' (实为 GLM-5.3)"] A --> B["技术/评估考量"] B --> B1["隐去中国开源模型标签"] B1 --> B2["获得全球开发者无偏见真实测试反馈"] A --> C["公关/博弈考量"] C --> C1["营造赛博朋克式神秘感引发自发讨论"] C --> C2["避免与美欧顶级巨头同日宣发直接撞车"] ``` Ash Minhas 从另一个视角分析了官方在揭晓时给出的理由：开发者隐去身份主要是为了获取全球开发者最纯粹、毫无偏见的真实反馈，以真实评判模型在各场景下的性能上限。Ash 认为，隐去“中国模型”这一标签确实在客观上排除了潜在的地域认知偏见，这一技术维度的考量大约占了 50%，但不可否认的是，剩下的 50% 绝对是一场教科书级别的营销奇谋。节目在轻松的告别氛围中落下帷幕，主持人 Tim Hwang 向所有嘉宾和听众致谢，并邀请大家下周继续收听《Mixture of Experts》。 - 突袭式发版避险：避免与 OpenAI、Anthropic 等前沿实验室正面交锋造成关注度被瞬间分流。 - 研发痛点共鸣：IBM Granite 4.2 刚发布即遭遇通义千问（Qwen）同体量强势新品压制，印证了宣发择期的不可控性。 - 真实反馈动机：去除品牌与地域标签能让开发者摒弃先入为主的刻板印象，专注于评估输出质量。 - 营销效果复盘：先匿名引发社区“寻宝式”自发解密，再高调宣布身份，其传播效率远超传统的新闻通稿。 ## Notable quotes > `27:30` 挑选在何时以及如何发布模型，本身就是一场高难度的投机博弈，你得去赌还有谁会在同一个空间、同一尺寸级别发布竞争模型。 — Gabe Goodhart > `00:00` This is a clear line in the sand from NVIDIA that they are going to depend, their future is going to depend on a strong, thriving, many model ecosystem. — Gabe Goodhart ## Mentioned Tim Hwang, Aili McConnon, Gabe Goodhart, Ash Minhas, Skyler Speakman, IBM, NVIDIA, Arm, Hugging Face, Poolside, Transformers, Llama.cpp, Nemotron, NemoClaw, Hermes, Open WebUI, IBM Z, LinuxONE, z/OS, PyTorch, OpenRouter, OX Alpha, GLM-5.3, Granite 4.2, Qwen, OpenAI, Anthropic, Hot Chips ## Video description On episode 122 of Mixture of Experts, hosts Tim Hwang and Aili McConnon are joined by Gabe Goodhart, Ash Minhas, Skyler Speakman to talk hot chips, mystery models, and major money moves. We open with IBM's Hot Chips conference reveal of its dual-architecture mainframe processor built with Arm that lets IBM Z and LinuxONE systems run Arm-native Linux workloads right alongside z/OS. We break down what this means for enterprises balancing legacy reliability with the fast-moving Arm software... [Source](https://cdnapisec.kaltura.com/p/1773841/sp/177384100/playManifest/entryId/1_rqeaunh4/format/download/protocol/https/flavorParamIds/0/name/media.mp3)

Source quotes

[27:30] Gabe Goodhart: 挑选在何时以及如何发布模型，本身就是一场高难度的投机博弈，你得去赌还有谁会在同一个空间、同一尺寸级别发布竞争模型。 [0:00] Gabe Goodhart: This is a clear line in the sand from NVIDIA that they are going to depend, their future is going to depend on a strong, thriving, many model ecosystem.

Source quotes

[5:00] Ash Minhas: Nvidia's play here is to get as many models trained and running on their CUDA cores. [5:00] Gabe Goodhart: the smaller hardware vendors that are trying to do commodity hardware that are not NVIDIA, they're trying to soak up excess demand, frankly.

Original transcript

[0:00] - This is a clear line in the sand from NVIDIA that they are going to depend, their future is going to depend on a strong, thriving, many model ecosystem. All that and more on today's mixture of experts. I'm Tim Huang, and welcome to Mixture of Experts. Each week, MOE brings together a panel of brilliant technologists working at the frontiers of artificial intelligence to lead you through the week's news. On this week's episode, we've got Gabe Goodhart, Chief Architect, AI Foundations, Ash Minhas, Technical Content Manager and AI Engineer, and Skylar Speakman, Senior Research Scientist at the Software Innovation Lab. Welcome to you all. I'm also joined by my co-host today, Like there's a lot of news this week, but this week it felt like there was maybe like an overwhelming amount of news. So we're going to cover a little bit, a new announcement about IBM's new next generation dual processor. We'll talk about OX Alpha and the kind of mystique around it, which just as of today actually was revealed who's behind it. But I actually want to start by talking a little bit about what's happening with NVIDIA, its recent line of acquisitions and why it's doing what it's doing. *music* The headline here is that Hugging Face announced that it was going to be acquired by NVIDIA for $12.9 billion. This follows on sort of just another deal not too long ago of a $6 billion kind of acquisition transaction with Poolside. And I guess, Gabe, you flagged both of these stories for us because you were just like... This is all about open. And so maybe I'll kick it over to you to give your hot take first. Yeah, I think this is a clear line in the sand from NVIDIA that they... are going to depend, their future is going to depend on a strong, thriving, many model ecosystem. And right now that is open. And right now they see that that is a critical lifeline for their company. So they're willing to put money behind it. They've clearly started down that track with their work on the Nemo Tron model line. And the poolside investment is clearly a further investment in making Nemo Tron a strong model line that people can rely on. And then this acquisition of Hugging Face is kind of just taking it to the max. Hugging Face is the cornerstone of the open AI ecosystem right now. And that's true of the models, but it's also true of the software. owned and operated by Hugging Face. It is the reference architecture of literally every LLM on the planet that is released in open source. So every downstream architectural implementation follows from the Transformers one. If I try to get something merged in Llama.cpp before it's merged in Hugging Fit in Transformers, they say, hold up, we can't count on this until that other PR is merged. So this is not, I mean, this is clearly a play for the center of gravity, right? So NVIDIA is making a clear statement here that they want to be the center of open source AI. And that, I think, has some... pros and some cons, right? So a big company like NVIDIA that is clearly well-moneyed, well-respected, and in a very strong position of power in this industry, it's great to have that support behind the open ecosystem. For those of us that live in that ecosystem, it means we've got backing for years, which is amazing. On the other hand, a big piece of the open ecosystem incentivizes making the model smaller and easier to run on a wide variety of hardware and putting that behind a vendor that is clearly biased in what hardware models should run on and the size and cost of that hardware creates a bit of a perverse incentive. So I'll be really curious to see how this plays out in terms of is it, you know, I think the obvious analogy that a lot of people are drawing is GitHub and Microsoft, where GitHub is the de facto standard for where code And thus far, it seems like GitHub and Microsoft have done a pretty decent job of keeping a firewall between the incentives of the closed source Microsoft ecosystem and the open source code ecosystem in GitHub. There's obviously product cross-pollination with Copilot, but they haven't. from the open source ethos of GitHub. So the hope here is that NVIDIA would do the same with Hugging Face and that Hugging Face would remain a shining light of open source ethos championship and openness and that, you know, NVIDIA would simply be putting their bet behind it. So we'll see what happens. I'd love to push more on what you've underlined. My question is who benefits more in this? Is it the open source community? [5:00] or NVIDIA, you know, what in the best case scenario, how do they each help each other? And, you know, what would be some of the concerns? And we can, you know, Skylar or Ash, if you want to jump in to join the conversation, please. Yeah. Remember when Microsoft bought GitHub? It's basically that again. Now, Microsoft done a very good job of how that acquisition's gone. I think that one of the interesting things with nvidia and and then acquiring uh hugging face is going to be that there's a lot of inference providers already providing inference to hugging face and so how will that work now are they reselling the same hardware that Nvidia providing them. There's going to be some interesting dynamics start to appear from that perspective. Open weights, open software, but closed hardware. Nvidia's play here is to get as many models trained and running on their CUDA cores. That's the larger draw here. So on the one hand, we do benefit from having easier access to a wide range of models with an asterisk as long as they run on Nvidia chips. So I think that's what's really going on here, both with the And with the model factory from poolside, they want to turn these models out as fast as possible because they're still in the business of selling hardware. Well, I think on that note, you know, it really leans into why this is important to NVIDIA and why NVIDIA is well positioned to be the one making this investment. Pretty much every other significant vendor in this space is in either the model space itself or the software space above the model. NeoClouds that are providing that inference. If you're looking at any of the big hyperscalers, their goal is to get you running. It doesn't particularly matter what the chip in the computer is. It matters what the API surface you're hitting is and what metering gateway you're going through. NVIDIA doesn't care. It cares about the chip under the hood. So there's no disincentive for Hugging Face to continue providing a wide range of inference providers as long as all those inference providers are buying chips from NVIDIA to power the hugging face. So I do think there's an incentive alignment here that's fairly unique to NVIDIA. Again, I think the big losers in this deal will be the uh companies trying to make commodity chips that are not nvidia right so you've got i think one of the articles i read about this clearly positioned this as a defensive play against all of the labs that are trying to do vertical integration and build their own chips because if the big labs stop needing to buy nvidia chips nvidia needs to sell their chips somewhere else and so they're really going to aim at that commodity hardware like go broad rather than deep type of uh sales that really lose out are the others attempting to sell broadly to, you know, general data center consumers. Yeah. Can I play bad guy here for a second? Because obviously there's some discomfort in the room with, you know, NVIDIA's nefarious plan here to keep everybody on their kind of closed hardware. Yeah. Is there another view here which is like, get real, right? Basically, these commodity players were never, ever going to catch up. NVIDIA was so far ahead. We're just recognizing reality, which is that the game has been lost in terms of certainly open hardware, but maybe even commodity hardware here, just given NVIDIA's extensive lead. Do people agree with that? This is in some ways realism... versus them really kind of closing off an avenue for commodity catch-up? I think, to me, yes. And I also think it's important to note that the smaller hardware vendors that are trying to do commodity hardware that are not NVIDIA, they're trying to soak up excess demand, frankly. Like, right now, we are in a very skewed supply-demand ecosystem. And so... this isn't going to change that dynamic. Now, if the demand really does drop off, that's where you're going to start to see this become a bit more of a food fight. And you're absolutely right. You know, NVIDIA is going to throw its weight around and, you know, maybe it's just calling a spade a spade. But I think right now, all of the possibility of other vendors losing out is just speculation because frankly you know if a data center needs chips uh they're probably going to look for nvidia first but if it's a you know, nine month waiting period, they might look elsewhere. And NVIDIA is probably okay with that because they're still selling as many chips as they can make. I think it's also interesting whether this is like a new, you know, taking this news, you mentioned, obviously, the licensing of the poolside model factory. Is this like a new, you know, bigger, broader playbook for AI consolidation? Because like, in that case, what struck me was, you know, they weren't acquiring it. They, you know, bought the license, they're hiring the team, you know, so it's sort of NVIDIA land in a variety of different kind of manners, but perhaps that's too simplistic. I think a key part of that poolside NVIDIA conversation was it's non-exclusive license. So NVIDIA is not necessarily doing a land grab there. They are going to be one of perhaps [10:00] tool companies that are going to be using pool, using poolside's approach to model creation. So that, that was kind of in the details there of that agreement. So yeah, we'll be interesting to see who else might be in line to pay $6 billion to poolside for their license. I don't think there'll be too many customers left in that. But this is not exclusive poolside and NVIDIA in that particular arrangement. Yeah. And that's maybe where I want to kind of like do the final few questions on this segment is, uh, Normally, when you spend $19, $20 billion, you're usually a little bit short on the money. But we're also talking about NVIDIA here where they could go for a lot more. I guess the question is, Ash, I'm curious if you have any thoughts on who's next up here? Where else would you go if you were trying to consolidate in the open space? Do you think NVIDIA is going to make another play? Or is this it for a little bit? I mean, the stack is like, you know, chips and infrastructure, then it's like the cloud providers providing that and then you've got the application layer. I think with both of these things, it's like infrastructure, lockdown, They're in that middle layer now and they want to make really good models. They're going to use, obviously, that poolside acquisition to make their own NemoTron models better. I'm curious. I know I'm responding to your question, Tim, with a question, but the question is, are they going to go even higher up the stack and start looking at some of that application layer and maybe going for a slice of that somewhere? And in making that open, or is it like going to go grab open web UI? They've already taken a small step in that direction with NemoClaw, right? So they're probably one of the largest companies behind an open source agent harness at this point. So, you know, that... There was a big announcement a little while ago about Nemo Claw, and it's kind of... I wouldn't say it's it's petered out, but there hasn't been as much hype about the clause as of late. I think a few other agent harnesses like Hermes have taken up a little bit more of the oxygen for the generalist agent. But, you know, I guess I would be surprised to see NVIDIA try to go the actual like for sale, you know, buy this as a consumer software platform route. we'll be going to like ai.nvidia.com and you know having our agent chats there the way folks do with chat gpt or claude it just doesn't seem like that incentivizes the right things because ultimately their business is still built completely around the chips that are under the hood so i think for me the two recent deals we're talking about here are just a clear sign that they are going for as many flowers blooming as possible right and if they can put their money and their mouth behind amplify that goal, they will. But I don't see them trying to then also carve off a corner of the market and say, but forget all those other flowers, come to our specific flower. I'm going to move us on to our next topic, which I guess is related in some ways, but it's a different side of the market entirely. We've talked about the mainframe market here on MOE in the past. And one of the things I love about the market is that it's very different from what we know of usual trends. you know, things around AI, where the general approach of the industry has been just like, all right, let's just launch it and see what happens. With mainframes, you really kind of can't afford that in some ways. And Ailee, I know you flagged this one, I don't know if you want to introduce it, but it sounds like IBM is basically out with a new dual processor in the mainframe space, which is a kind of big deal, just given, you know, how cautious people are in mainframes. Do you want to talk more about it? Sure, sure. And yeah, they presented this at this week was chip makers, researchers come together, unveil new processors and designs. And, you know, many of the participants were, you know, some of those, Gabe, that you mentioned, the frontier companies who've all come up with their own kind of custom chips that they're sharing. But in this context, IBM was announcing kind of a different type of chip, you know, this dual process architecture. And so this one they've developed with ARM, the AI software ecosystem. And what's different about it is it IBM's Z, so the mainframe workload with ARM and all on one chip and with the ability for the ultimately to be able to switch between those two workloads, you know, in nanoseconds. So I think the bigger picture, you know, thinking is that there's a growing challenge in enterprise AI where organizations, you know, want to run their inference closer to where that data resides. systems. [15:00] And at the same time, you have this ARM ecosystem where a large share of the modern AI software is being built. So I think that, you know, the idea is rather than moving the data between these two places, IBM is sort of, you know, proposing this chip to bring those workloads closer together. And maybe, Skylar, if you're game to, you know, jump in on this first, I'm interested, you know, from an infrastructure perspective, why does this matter? You know, how big of a deal is this? details about how these processors work, there's two different philosophies. One of them is doing multiple simpler instructions. So if you want to add two numbers, there's one instruction to go read a number from memory, there's another instruction to add it, and there's another instruction to write it. The other philosophy is a single, more complex instruction that does that all end to end. And so computer science over decades has evolved from kind of those two these worlds somewhat collide in the IBM arm space. I do not know the technology behind it about how they're doing both of those instruction sets on the same chip. Kudos to them. Very smart people working on that. But it really, I think it's going to be interesting to ask questions like, is your bank going to go to the app store and update its core banking software? Can you go to an ATM and do a transaction and say, wait a minute, our backend is training model right now, give us 20 seconds. And those are questions we had asked before, because these things have been completely separate ecosystems. And now IBM and Arm are saying, still mainframe functionality, but the ability to run these two different instruction sets. And they have, again, what I want to emphasize here is two different philosophies of how you go about writing code coming together under one place. And the thing that kept bothering me when I was reading this is On the one hand, it's cool to see these fences being brought down, but there's also this approach where you should be asking, why were those fences there to begin with? Do we really want to have ARM software sitting in the same places as, for example, our government services, banking services, and the things that have relied on mainframes for decades? So watch this space. I don't know who else can comment on that. instruction sets. It really is two different worlds colliding, and it'll be really cool to see it play out in the coming months. And I think it's sort of, yeah, I think to your point of it being a larger question, too, it's sort of this looking at enterprises, trying to integrate, you know, AI with these mission-critical systems, these, you know, established systems that, you know, they can't just totally rip out and then, you know, kind of bring in, you know, something totally new. And the analogy you used of the App Store, someone, as I was talking about this news, It's like mainframe consumers have this whole app store of options if they're an enterprise, hoping to bring in the inference kind of closer to where their data is sitting, and it's harder to get out of. I think both IBM and ARM are winners in this case as of now. I think right now, I don't think this is a zero-sum game here. I think there's opportunities for both of these kind of established players. And then a few months later, the announcement of the chip will still have to wait for its actual release. Maybe maybe I'm a bit doubtful for that, but it's cool to see that sort of quick turnaround on this type of partnership. Yeah. So I want to bring in the perspective of a software engineer on this, which is this is going to make a lot of people's lives a whole lot easier. And the market worried that that meant the death of the mainframe. And we had a lot of well-articulated responses out of IBM. But I think the core of all those responses was, Look, COBOL is a means to an end, but the end has not gone away. Right. The reason mainframes exist is bulletproof reliability. Like I was talking to I forget who it was earlier this week, who was literally part of the team that said you could. literally take the mainframe out back and put a bullet through it and it would not stop processing transactions. And that's just not true of any other piece of hardware, right? Like you can, you could shoot a bullet through a RAM stick on a Z system and it wouldn't blink. Now that reliability is what the foundation of many of our most critical pieces of infrastructure is built on. And so, yes, there's a lot of code trapped in COBOL, which is a pain in the neck, and is probably very difficult for companies to find programmers that can actually handle that code. [20:00] the need for that reliability has gone nowhere. And so this is an attempt to say, well, we can solve that problem in a different way. Rather than bringing the mainframe code out and then sacrificing that reliability that you've come to depend on, why don't we bring the... modern code in, right? And as someone who has tried to cross-compile code for arbitrary random architectures, if you're thinking about import Torch as a Python program and you're like, what does cross-compiling mean? Well, let me tell you, Python is a C program. Every single library that you import in Python is in fact delegating down to a C library if it's got any kind of architecture out there. So it is a massive pain in the neck of, you know, ecosystem targeted cross compilation. And the idea before this announcement of taking something as complicated as an LLM, that is backed by all of this complicated acceleration logic and cross compiling it all to a completely bespoke, bespoke is the wrong word, completely firewall and separate chip And this was the purview of, you know, potentially years of work to get, you know, individual LLM architectures ported over to a Z system. Now, with ARM and the fact that a huge number of people are running ARM workloads on their laptops, on their phones, even starting to be on their desktop processors and server processors, means that a huge amount of software has already been adapted to the ARM ecosystem. So that work is done for us. going to open up the floodgates of bringing a ton of interesting workloads to the mainframe that just could not run there in any reasonable amount of effort before. Now, to your point, Skyler, it'll be interesting to see... what that does for the reliability of the other workloads that are sharing the processor. So the proof will be in the pudding there. But at least on the surface of it, this looks like this is going to make a lot of people's lives a lot easier. Yeah. And can we talk a little bit about the future, Ash? Like I'm thinking about like in 2050. Will we still have mainframes? It feels like this technology that people really gripe about and are always like, the mainframe's on its way out, the mainframe's on its way out. But to Gabe's point, it is rare to find reliability of this kind. And, you know, I guess the question is, I guess in some ways you could almost read the trend line. It's like actually mainframes will be here and sort of bigger than ever. But do you buy that? I guess I'm kind of curious about like whether or not this like pretty dusty kind of ecosystem in some ways is like much, much more robust and interesting than it looks. I think that there are certain workloads where latency really makes a difference. OK, and the latency. Yeah. issue just doesn't really exist in mainframe. You know, you're getting transactions processed in such a fast time that, uh, we will always need the ability to do that with certain types of transactions. And so I think that we will trend towards sort of, bigger, more powerful, computers of this kind. pretty much forever, really. The point that Skylar and Booth gave me, I also have some questions, I would say, right now, and curiosity. This is a great sort of thing that's come out of this collaboration with Arm. And I guess... the when i think about this from sort of like a computer architecture perspective they could have just put some arm cores into the chip and they did it okay it's all combined in one and that's really really cool and really really interesting i would love to like go and talk to the people who made that architectural decision and go why did you decide to go the hard way because it could have been easier if you did it the other way um and yeah also that then raises questions of like well you know we're trying to process these things on mainframe today, which are like sort of millisecond type transactions. And so now if you've got, you know, I don't know, let's say PyTorch running on there and something else, okay, what does that do to the behavior of the overall system? And is that going to mean that it will need a much, much bigger computer in 2050? Yeah, that'll be a really funny outcome. It's just like we're living in 2050. And it's like, giant mainframes. That'd be really interesting to see. I'm going to move us on to our kind of last topic of the day. You know, I think it's an adage in 2026 that it feels increasingly like we're living in a cyberpunk novel of some kind. This story, I think, was very much like it. Basically, on Open Router, there's this mystery model called OX Alpha that kind of hit. And people are very impressed by its capability. And most interestingly, you know, at least in the announcement, it was claimed that it was able to serve 100 trillion tokens a day. [25:00] and, you know, which immediately raised questions about like, what kind of compute are you running to be able to offer this kind of thing? I think as of this morning, it has finally been revealed that it is GLM 5.3, open source Chinese model. And I guess maybe the first thing is we should just do the usual vibe check test. Gabe, Skyler, Ash, I'm curious if any of you kind of played with the model. What do you think about it just from a, you know, almost kind of like a wine and one of these things that I think in this like... probabilistic world are going to just be contested forever. So at some point, I don't know, about a year ago, I just put together sort of like, you know, my own little like benchmark suite, whatever. And I ran that through Open Router with it. And I thought, you know what? This is actually pretty good. pretty good i mean it's great that it's free it's not free now but it's still dirt cheap and um i was pretty pretty impressed with with the output that i got yeah i mean i gave it the smell test as well um it smells like a good model um but like with wine i at this point can't actually claim a connoisseur's palette uh because most of the workloads i want to run can be satisfied by a sub 30 billion parameter model on my dgx sparks so thank goodness for that yeah so you know little plug for small models there, but my feeling is that It is fantastic to have competition at the top end. Love seeing it. Very curious whether this throughput number that they are claiming is due to breadth of scaling and just raw compute, or whether they've done something truly novel in the model architecture that enables massive throughput improvements while maintaining this top end quality. That would be genuinely very cool if they figured some tricks out about how to, you know, get the bits through the attention mechanism faster. But at the end of the day, we have a glut of very, very, very good models. And that's awesome. And I'm still going to try to run as much as I can off my local machine. My interest in this story is as much about how they decide to go secret. And, you know, I'm interested in your thoughts on, you know, presumably intentional. Does it build hype? You know, is there more to it than that? field at some point? Why the stealth mode? I think the stealth mode worked. I mean, in between the time where we had these topics chosen for this uh for this podcast and when it's actually recorded it came out as glm 5.3 would we be talking about the release of glm 5.3 if that was just how it came about or is it so much more fascinating to have this idea of a hidden model as tim pointed out it's like a cyberpunk novel where you can have these you know mysterious people showing up and competing in a i don't know a medieval joust with uh with helmets on we don't know who they are um and so i think the makers of the model to release it in this way. They got the hype they wanted to. A few days of great speculation, and then they released saying it is just an improvement, not just. It is an impressive improvement over one of their previous models. So I think they really did a great job with the hidden reveal, letting the internet talk about it for two or three days. And I will say another thing to that, which is that timing a model release is very hard. We at IBM just released the Granite 4.2 models, which, you know, We are proud of they are certainly not going to win the benchmark race, but we think that they're going to be very useful for the target audience. However, a day after we did that, we had Quinn drop yet another benchmark busting amazing model that also runs on the same local hardware. And, you know, picking when you want to release and how you want to release is a great. a very difficult game of speculation on who else is going to be releasing competitive models in the same size, in the same space. Um, so, so like you said, Skylar, I mean, kudos to the team for choosing a, um, different route to release because if this had been just another amazing Chinese model, like what a world we live in, by the way, you know, would we even bother to talk about it? But, you know, choosing a release strategy that almost was competition proof because nobody else was doing it. Now, I don't think anyone else can probably pull this off again for a little while. I mean, we had it kind of with Nano Banana for a while. This isn't the first time a lab has sort of stealth launched a model with a catchy name, but... Um, It is a good strategy to mitigate against, oh, shoot, is one of the other Frontier Labs also going to release their new model on the same day or the same week? Who knows? Or sometimes within, you know, I'm recalling one OpenAI Anthropic pair of releases within like an hour of each other where it's sort of it becomes the sort of playground battle of who gets most attention. I think that they did have like a like an announcement to say, oh, basically. [30:00] wanted to release it in a sort of stealth mode so they could get unbiased actual real feedback from developers to to know you know how good their model is in all of those use cases i do think there's probably an element of that in like why it was stealth right you know they didn't want to say it's a chinese model and for it to like introduce any sort of bias into determination as to the the model's uh performance um i i think i'm going to say that's a little bit of a little bit I do also think there's probably 50% of it is marketing, right? It's a great marketing thing to do. That's great. Well, that's all the time that we have for today. Gabe, Sky, Ash, it was great to have you on the show. And I was great co-hosting with you today. Thank you, Tim. And thanks for joining all you listeners. If you enjoyed what you heard, you can get us on Apple Podcasts, Spotify and podcast platforms everywhere. And we'll see you all next week on Mixture of Experts. *music*