Multiverse Prunes AI Models to Cut Compute Costs Up to 80%

Sovereign and Efficient AI: The Case for Optimized AI Models in the Age of the Inference Crunch — Ai4

Ai4

Enterprise software margins face sustained destruction as compute scarcity and rising agentic inference outpace hardware efficiency gains over the next decade.

AI-assisted summary of Sovereign and Efficient AI: The Case for Optimized AI Models in the Age of the Inference Crunch — original source. Check the original for full context.

Source published: 2026-08-17

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 17, 2026, this presentation features Chris Zaharias, VP of U.S. Operations at Multiverse Computing. According to the page description, the talk addresses the enterprise AI ROI crisis caused by sold-out GPU capacity and escalating inference costs, demonstrating how Multiverse Computing's sovereign and optimized AI models can cut inference expenses by 50–80%.

## TL;DR Chris Zaharias argues that persistent compute scarcity and eroding profit margins are driving an enterprise 'inference crunch' that technological advancements alone cannot resolve due to Jevons' paradox. He presents Multiverse Computing's quantum-based tensor network pruning, a technique that compresses large AI models by 50% to 60% with minimal accuracy loss—and occasionally improved benchmark performance. By enabling efficient on-device deployment across millions of AI PCs, edge hardware, and standard CPUs alongside intelligent query routing, Multiverse provides enterprises with a clear path to drastically lower cloud inference costs and achieve AI independence.

## Atmosphere & style The presentation carries a fast-paced, authoritative, and analytical tone reminiscent of a 'Mary Meeker style' market deck presentation. Zaharias delivers a dense, data-driven narrative with urgent realism about software margin destruction, balanced by confident enthusiasm for mathematical innovation. Watching the session feels like attending a high-stakes executive briefing where complex subatomic physics meets practical enterprise cost engineering.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=4s) The primary threat to enterprise AI ROI is infrastructure economics and skyrocketing compute costs, which are rapidly creating a severe financial and environmental hazard. - [00:04](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=4s) Layering heavy AI inference onto existing platforms is destroying gross margins and terminal enterprise value for software vendors. - [00:04](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=4s) Due to massive exponential demand, the severe compute and hardware supply crunch will persist for 5 to 10 years rather than resolving quickly. - [00:04](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=4s) Jevons' paradox will override Moore's law efficiency gains, as cheaper inference causes overall enterprise demand to outpace hardware advancements. - [00:04](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=4s) Multiverse Computing uses mathematical algorithms derived from quantum physics research to map and prune redundant, unused neural network parameters. - [00:04](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=4s) Unlike standard quantization, quantum-based tensor network pruning performs intelligent structural reductions while preserving baseline model accuracy. - [00:04](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=4s) Multiverse's compression pipeline reduces model parameter count by 50% to 60% and memory footprint by ~70% with only 2% to 3% accuracy loss. - [00:04](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=4s) Through a partnership with a major PC manufacturer, a 20-billion parameter model was compressed to 6 GB, enabling on-device AI across 40 to 50 million PCs. - [08:00](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=480s) In some cases, compressed models outperform their bloated, uncompressed baselines across public benchmarks by removing structural inefficiencies. - [08:00](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=480s) Major enterprises are seeing cloud compute bills for LLMs escalate exponentially from 7-figure amounts ($1M+) to 9-figure sums ($100M+). - [08:00](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=480s) Multiverse provides model-agnostic optimization across 150+ models, delivering solutions via software licensing for local deployment or Inference-as-a-Service. - [12:00](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=720s) Future AI inference is shifting away from centralized cloud data centers toward local edge deployment on end-user hardware, laptops, and robotics. - [12:00](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=720s) Automated query orchestration layers like the Capacify Router evaluate prompt complexity to direct requests to the most cost-effective model, cutting inference costs by 50%+. - [12:00](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=720s) Sufficiently compressed AI models will allow the majority of future enterprise inference to run on standard CPUs, NPUs, and alternative hardware rather than expensive GPUs.

## Detailed walkthrough Here is a detailed, faithful walkthrough of the transcript:

[00:04](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=4s) The Software Margin Squeeze and the Impending Inference Crunch

Chris Zaharias, VP of U.S. Operations at Multiverse Computing, opens his talk in a fast-paced "Mary Meeker style," aiming to move through a dense deck of slides while providing broad industry context. He introduces Multiverse Computing as a 500-person Spanish AI firm that recently achieved unicorn status after raising around $500 million, driven by demand for inference efficiency. The company’s primary thesis, formed years earlier at the same conference, posits that training and compute costs are skyrocketing into a "stratospheric" environmental and financial hazard. While this initial squeeze is currently destabilizing the software sector, Zaharias warns it will inevitably affect every industry over the next 5 to 10 years as inference becomes as fundamental to business operations as hiring employees or running marketing campaigns.

Focusing on the independent software vendor (ISV) landscape, Zaharias highlights three converging trends that threaten enterprise ROI: * **Eroding Profit Margins:** Over the past year, market capitalizations for top publicly traded software companies have plummeted because layering AI inference onto existing platforms rapidly shrinks gross margins and destroys terminal enterprise value. * **Durable Compute Scarcity:** Contrary to optimistic forecasts that memory, GPU, and CPU shortages will resolve within a year or two, Zaharias argues the supply crunch will persist for 5 to 10 years due to exponential inference growth. * **Jevons’ Paradox vs. Moore’s Law:** While technological advancements continue to lower token costs, Jevons' paradox will override these gains—as inference becomes cheaper and more accessible, aggregate demand across enterprises will massively outpace hardware efficiency gains.

Zaharias emphasizes that CFOs at tech conferences are struggling to present credible solutions to investors regarding AI-driven margin destruction, leaving stock valuations vulnerable across the sector.

---

[00:04](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=4s) Quantum Math Roots and Intelligent Neural Network Pruning

Originating from the Basque region of Spain—which Zaharias affectionately compares to "Spain's version of San Diego or La Jolla"—Multiverse Computing leveraged long-standing government funding in fundamental quantum physics research. The company's four co-founders developed novel mathematical algorithms originally designed to model quantum many-body states (highly intercorrelated subatomic particles). They discovered that this specific mathematical framework is exceptionally effective at mapping, profiling, and removing redundant, unused parameters within deep learning models, including LLMs, vision-language models, and computer vision architectures.

Contrasting this breakthrough with traditional compression techniques like standard quantization (which uniformly reduces numerical precision or cuts neurons at the expense of performance), Multiverse performs "intelligent pruning." By analyzing the complex connectivity between individual neurons, the algorithm identifies which connections genuinely contribute to the network's output and strips away useless weights. Because modern foundation models are wildly overparameterized, this approach allows for significant structural reductions without triggering an accuracy drop:

``` [ Uncompressed Large Model ] │ ▼ ┌──────────────────────────┐ │ Quantum-Based Profiling │ ──► Analyzes neuron connectivity └──────────────────────────┘ │ ▼ ┌──────────────────────────┐ │ Tensor Network Pruning │ ──► Strips unneeded weights/parameters └──────────────────────────┘ │ ▼ [ 50–60%+ Compressed Model ] ──► Retains baseline accuracy ```

* **Core Compression Performance:** Reliably reduces model parameter count by 50% to 60% (and up to 95% in extreme cases) with a ~70% reduction in memory footprint and minimal accuracy loss (typically 2% to 3%). * **PC Deployment Milestone:** A major PC manufacturer partnered with Multiverse to achieve on-device AI across 50 million annual units. Multiverse successfully compressed a 20-billion parameter model to fit within a 6 GB memory budget on a 16 GB "AI PC," leading to direct investment from the manufacturer and planned shipments on 40 to 50 million PCs over the next year.

---

[00:08](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=8s) Outperforming Baseline Models and Real-World Enterprise Impact

Zaharias reveals a counterintuitive outcome of their quantum compression pipeline: in certain cases, compressed models actually outstrip their larger, uncompressed baselines. For instance, Multiverse's compressed version of GLM 5.2 uses 50% fewer parameters and 70% less memory, yet it outperformed the raw GLM 5.2 model across three out of six public coding benchmarks. Zaharias attributes this to the fact that venture-backed foundation model developers prioritize brute-force scale over efficiency, building bloated models to maximize token consumption.

This inefficiency is causing staggering enterprise expenditure. Zaharias notes that over a three-month span, he met with numerous publicly traded companies whose annual compute bills for Anthropic (Claude) and OpenAI escalated from seven-figure figures ($1M+) to nine-figure sums ($100M+).

To combat this, Multiverse operates a flexible model-agnostic business model: * **Broad Architecture Support:** They have compressed over 150 models to date, including Qwen, GPT-OSS, DeepSeek, and YOLO for computer vision. * **Proprietary Model Optimizations:** In a recent project for an enterprise software client managing over 100 proprietary models, Multiverse compressed their embedding, speech-to-text, text-to-speech, and vision-language models by 50% to 60%. This allowed the client to offload heavy cloud compute to local devices, curbing an aggregate cloud inference bill that was projecting a jump from $500 million to $1 billion. * **Flexible Delivery Models:** Offerings are available via direct software model licensing for local/on-prem deployment or through a reduced-cost Inference-as-a-Service API hosted on lighter AWS infrastructure.

---

[01:12](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=72s) Edge Deployment, Intelligent Routing, and the Shift to CPUs

Looking forward, Zaharias asserts that the future of enterprise inference lies outside centralized cloud data centers. The industry is rapidly shifting toward running local inference directly on end-user hardware, such as laptops, mobile devices, set-top boxes, and robotics. To accelerate this transition, Multiverse has partnered with a major global edge infrastructure provider (historically known for pioneering edge content caching via proxy servers) to host compressed models across a distributed edge inference cloud.

To further manage operational expenses, Zaharias stresses the necessity of intelligent request routing. Multiverse offers the "Capacify Router," an orchestration layer that evaluates incoming prompts and automatically directs them to the most cost-effective model suited for the task. Rather than routing simple queries to expensive tier-one models like GPT-4 or Claude, intelligent routing matches the prompt's requirements against a spectrum of speed, cost, and generation quality:

* **Cost Reductions:** Implementing an automated query router can instantly cut enterprise inference expenses by over 50%. * **Hardware Independence:** Zaharias predicts that the majority of future AI inference will shift away from high-end Nvidia GPUs toward existing enterprise CPU architectures, NPUs, and hardware from Intel, Qualcomm, and AMD. With sufficient model compression, standard x86 and ARM processors can handle heavy production workloads at a fraction of the cost.

Closing his presentation, Zaharias invites attendees to visit the Multiverse Computing booth to inspect live demonstrations of compressed models running on edge hardware and explore custom enterprise integration strategies.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=0s) The biggest threat to enterprise AI ROI isn't model quality it's infrastructure economics. — Chris Zaharias

> [00:00](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=0s) With GPU capacity sold out, hyperscalers rationing access, and agentic workflows multiplying compute consumption 10–100x, companies are hitting a wall. — Chris Zaharias

> [00:00](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=0s) The answer isn't waiting for new chips it's sovereign, efficient AI: compressed models that run where you need them, at the scale you can afford. — Chris Zaharias

> [04:00](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=240s) Over the past year, market capitalizations for top publicly traded software companies have plummeted because layering AI inference onto existing platforms rapidly shrinks gross margins and destroys terminal enterprise value. — Chris Zaharias

> [04:00](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=240s) Contrary to optimistic forecasts that memory, GPU, and CPU shortages will resolve within a year or two, Zaharias argues the supply crunch will persist for 5 to 10 years due to exponential inference growth. — Chris Zaharias

> [04:00](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=240s) While technological advancements continue to lower token costs, Jevons' paradox will override these gains—as inference becomes cheaper and more accessible, aggregate demand across enterprises will massively outpace hardware efficiency gains. — Chris Zaharias

> [04:00](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=240s) By analyzing the complex connectivity between individual neurons, the algorithm identifies which connections genuinely contribute to the network's output and strips away useless weights. — Chris Zaharias

> [04:00](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=240s) Reliably reduces model parameter count by 50% to 60% (and up to 95% in extreme cases) with a ~70% reduction in memory footprint and minimal accuracy loss (typically 2% to 3%). — Chris Zaharias

> [08:00](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=480s) In a recent project for an enterprise software client managing over 100 proprietary models, Multiverse compressed their embedding, speech-to-text, text-to-speech, and vision-language models by 50% to 60%. — Chris Zaharias

> [01:12:00](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=4320s) Implementing an automated query router can instantly cut enterprise inference expenses by over 50%. — Chris Zaharias

> [01:12:00](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=4320s) Zaharias predicts that the majority of future AI inference will shift away from high-end Nvidia GPUs toward existing enterprise CPU architectures, NPUs, and hardware from Intel, Qualcomm, and AMD. — Chris Zaharias

## Mentioned Chris Zaharias, Multiverse Computing, Ai4, Nvidia, Intel, Qualcomm, AMD, Anthropic, Claude, OpenAI, GPT-4, GPT-OSS, Qwen, DeepSeek, YOLO, GLM 5.2, AWS, Capacify Router

## Video description Chris Zaharias - VP U.S. Operations, Multiverse Computing

The biggest threat to enterprise AI ROI isn't model quality it's infrastructure economics. With GPU capacity sold out, hyperscalers rationing access, and agentic workflows multiplying compute consumption 10–100x, companies are hitting a wall. The answer isn't waiting for new chips it's sovereign, efficient AI: compressed models that run where you need them, at the scale you can afford. Multiverse Computing will show how efficient and optimized AI models that reduce their inference costs by 50-80%, giving enterprises, ISVs, and governments a concrete path to AI independence and cost control.

[Source](https://www.youtube.com/watch?v=h-e2XJUgZ3o)

中文简报

Multiverse剪枝AI模型，算力成本最高削减80% 未来十年算力匮乏与智能体推理需求的激增将超过硬件效率的提升，企业软件利润率面临持续冲击。 Chris Zaharias指出，持续的算力匮乏与被侵蚀的利润率正在引发企业的“推理危机”，而受杰文斯悖论影响，仅靠技术进步无法解决这一问题。他介绍了Multiverse Computing基于量子的张量网络剪枝技术，该技术能在精度损失极小（甚至偶尔能提升基准表现）的情况下将大型AI模型压缩50%至60%。通过在数以百万计的AI PC、边缘硬件和标准CPU上实现高效端侧部署，并配合智能查询路由，Multiverse为企业提供了一条大幅降低云端推理成本并实现AI自主的清晰路径。企业 AI 投资回报率（ROI）面临的最大威胁已转向基础设施经济学与算力短缺，预计算力紧缺将持续 5 至 10 年，且杰文斯悖论将使总体推理需求增速远超硬件效率的提升。 (0:04) Multiverse Computing 利用源自量子物理研究的张量网络算法对神经网络进行结构化智能剪枝，可在参数量减少 50% 至 60%、内存占用减少约 70% 的同时，将精度损失控制在 2% 至 3% 以内。 (0:04) Multiverse 成功将一个 200 亿参数（20B）的模型压缩至 6 GB 内存预算内，并与头部 PC 厂商合作，计划在未来一年内装配于 4000 万至 5000 万台端侧 AI PC。 (0:08) 为应对企业级 AI 推理账单从数百万飙升至上亿美元的压力，采用 Capacify Router 等智能路由结合向边缘设备及常规 CPU 架构的迁移，能够直接削减超过 50% 的推理成本。 (1:12)

Reading summary

Chris Zaharias of Multiverse Computing argues that escalating AI inference costs threaten enterprise margins, requiring quantum-inspired model compression and edge deployment to replace brute-force cloud compute. The Enterprise Margin Squeeze and Compute Scarcity (0:04) According to Zaharias, the deployment of large language models is eroding gross margins across the independent software vendor sector. While token costs continue to decrease, Zaharias asserts that Jevons’ paradox will drive aggregate demand far past hardware efficiency gains, causing annual vendor compute bills with providers like Anthropic and OpenAI to escalate from seven-figure to nine-figure sums. Contrary to expectations of a near-term resolution, he forecasts that shortages across GPUs, CPUs, and memory will persist for 5 to 10 years as inference becomes core to everyday business operations. Quantum Mathematical Algorithms for Model Pruning (0:04) Originating from quantum physics research in Spain's Basque region, Multiverse Computing uses mathematical frameworks initially developed to model quantum many-body states to prune neural networks. Unlike standard uniform quantization, which cuts numerical precision at the expense of model quality, this method analyzes inter-neuron connectivity to remove redundant weights from overparameterized foundation architectures. Zaharias reports that the technique reliably cuts model parameter counts by 50% to 60%—and up to 95% in extreme instances—while reducing memory footprint by roughly 70% with an accuracy loss typically held to 2% to 3%. In a production deployment with a major PC manufacturer, Multiverse compressed a 20-billion parameter model to operate within a 6 GB memory budget on a 16 GB PC, supporting planned deployments across 40 to 50 million units over the following year. Efficiency Gains and Architectural Flexibility (0:08) Zaharias argues that venture-backed foundation model creators intentionally prioritize scale over efficiency to increase token consumption. As a consequence, pruned models can occasionally match or exceed baseline performance; for example, a compressed variant of GLM 5.2 with 50% fewer parameters and 70% less memory outperformed its uncompressed counterpart on three out of six public coding benchmarks. Multiverse has applied this pipeline to over 150 architectures, including DeepSeek, Qwen, GPT-OSS, and YOLO. In an enterprise engagement spanning more than 100 proprietary speech, vision-language, and embedding models, compression of 50% to 60% allowed workloads to migrate locally, arresting an enterprise cloud bill on track to climb from $500 million to $1 billion. Decentralized Edge Inference and Hardware Shifts (1:12) Zaharias projects that long-term AI inference will shift away from centralized cloud data centers toward localized edge hardware, such as personal computers, mobile devices, set-top boxes, and robotics. Multiverse is deploying compressed models onto a distributed edge inference cloud in partnership with a global edge infrastructure provider. To complement on-device execution, Zaharias advocates for intelligent query routing via systems like the Capacify Router, which evaluates prompt difficulty to direct queries to cost-appropriate models rather than defaulting to frontier architectures like GPT-4 or Claude. He claims automated routing can cut enterprise inference expenses by over 50%, while forecasting that the majority of future enterprise inference will transition from high-end Nvidia GPUs to standard enterprise CPUs and NPUs from Intel, Qualcomm, and AMD. Zaharias concludes that resolving the enterprise inference crisis requires moving away from brute-force cloud GPU scale in favor of mathematical model compression, intelligent query triage, and execution on existing edge and CPU hardware.

精炼阅读摘要

Multiverse Computing高管Chris Zaharias指出，生成式AI推理正在急剧侵蚀软件企业的利润空间并面临长达5至10年的算力紧缺；企业亟需通过源自量子物理数学的算法剪枝、端侧迁移以及智能路由来大幅降低模型运行成本。软件毛利挤压与长期的算力短缺危机 (0:04) Zaharias认为，将AI推理层叠到现有软件平台上正在迅速压缩独立软件开发商（ISV）的毛利率，甚至损害企业的长期终值估值。尽管有乐观预测认为GPU和内存短缺会在一到两年内缓解，但他指出，在企业推理需求爆炸性增长的推动下，算力短缺将持续5至10年。此外，根据杰文斯悖论（Jevons' Paradox），随着模型调用单价下降，整体业务需求量的膨胀速度将远超硬件效率的提升幅度。算力成本失控直接反映在企业的财务报表上。Zaharias在走访中发现，多家上市公司在Anthropic和OpenAI上的年度计算支出已在数月内从百万元级跃升至上亿美元级；某家运营上百个专有模型的企业，其云端推理账单甚至面临从5亿美元攀升至10亿美元的压力，使得企业CFO普遍难以向投资者提供合理的毛利保障方案。量子算法剪枝打破模型超参数冗余 (0:04) 针对目前基础模型普遍过度参数化且资源利用低效的现状，Multiverse Computing利用其在西班牙巴斯克地区积累的量子物理研究成果开发出优化算法。该方法脱胎于原本用于模拟量子多体状态的数学工具，与简单压缩数值精度的传统量化技术不同，它通过精细测绘神经元之间的连接重要性，剔除对输出无实质贡献的冗余权重，实现结构性智能剪枝。根据该机构披露的数据，该算法能够稳定将模型参数减少50%至60%（极端案例下可削减95%），降低约70%的内存占用，同时通常仅带来2%至3%的精度折损。Zaharias以压缩版GLM 5.2为例指出，在减少一半参数和70%内存后，该模型在六项公开编程评测中有三项甚至超越了未压缩的原始基准模型。目前，某主要PC厂商已向该公司投资，并计划在未来一年内出货的4000万至5000万台AI PC上，将一个200亿参数模型压缩并部署至6 GB的内存预算内。端侧部署与硬件算力向CPU转移 (1:12) Zaharias断言，未来企业推理的重心将逐步离开集中的云端数据中心，向笔记本电脑、移动终端、机顶盒和机器人等边缘硬件迁移。为此，Multiverse已与一家全球边缘基础设施提供商达成合作，利用其分布式边缘节点网络托管和运行压缩模型，以降低通信延迟与云端依赖。在硬件路线上，Zaharias预测推理负载将逐步打破对昂贵Nvidia GPU的依赖，转向现有的CPU与NPU计算架构。他认为，只要模型经过深度的结构优化与参数剪枝，包括Intel、AMD、Qualcomm旗下的标准x86与ARM处理器，即可以远低于专用GPU的运行成本承担起实际的生产级推理负载。动态请求路由与灵活服务交付 (1:12) 除直接压缩模型外，Multiverse还强调通过智能分流来控制日常运营开支。该公司推出的“Capacify Router”编排系统能够实时分析用户输入的Prompt特征，并在速度、成本与生成质量之间进行权衡，自动将任务派发至最具性价比的适用模型，避免将所有简单任务无差别提交给GPT-4或Claude等昂贵模型。据估算，仅引入这一路由机制便可直接将企业的推理费用削减50%以上。在商业落地模式上，Multiverse采取中立的模型适配策略，已对Qwen、DeepSeek、GPT-OSS以及计算机视觉领域的YOLO等150多个架构进行剪枝优化。企业客户既可以通过纯软件授权的方式在本地端侧运行，也可以调用基于轻量化AWS基础设施的推理即服务（Inference-as-a-Service）API，以灵活控制混合架构下的总体拥有成本。 Multiverse Computing提出了以量子剪枝算法和本地化部署破解软件利润危机的技术方案，但该模式能否在主流企业中完全替代云端高规格集群，仍取决于边缘芯片生态的适配进度以及高度剪枝模型在复杂生产任务中的长期鲁棒性。

Detailed notes — zh

## Background 该视频由行业会议频道 Ai4 于 2026 年 8 月 17 日发布在 YouTube 上，演讲嘉宾为 Multiverse Computing 美国区运营副总裁 Chris Zaharias。内容简介指出，当前企业 AI 投资回报率的最大威胁并非模型质量，而是基础设施经济学与 GPU 算力配额限制，Multiverse Computing 旨在通过高效压缩与优化技术将推理成本降低 50% 至 80%，助力企业实现 AI 自主可控与成本控制。 ## TL;DR 演讲指出，由于杰文斯悖论与算力紧缺持续，高昂的推理成本正在严重侵蚀软件企业的利润与估值。Multiverse Computing 运用源自量子多体物理的张量网络数学体系对大模型进行精准剪枝，在保持或提升精度的同时实现 50% 到 70% 的极限压缩，成功将 20B 模型装进仅需 6GB 内存的端侧 AI PC 中。通过模型授权、低成本推理 API 与智能路由分发，企业能够将数亿美元级别的算力账单直接腰斩。最后，演讲大胆预言推理的终局并非英伟达的高端 GPU 垄断，而是重回存量庞大且成本低廉的通用 CPU 与端侧 NPU 时代。 ## Atmosphere & style 整场演讲语调务实干练、直击行业痛点，讲者以自嘲非技术背景的幽默化解了量子数学的晦涩感，叙事举重若轻。从开篇揭露上市公司面临“九位数算力账单”与毛利率崩塌的行业焦虑，到随后展示端侧极限压缩实测数据的从容自信，情绪曲线由危机警示转向破局突围。整体氛围充满商业洞察力与颠覆主流硬件认知的大胆底气，带给观众一种既冷静透彻又令人振奋的技术落地体验。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=4s) AI 推理成本的指数级飙升正在侵蚀软件企业（ISV）的毛利率，导致投资者下调对其终值预期并引发市值暴跌。 - [00:04](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=4s) 由于杰文斯悖论（Jevons Paradox）的作用，推理单价下降激发了更大规模的 Token 消耗，导致 GPU 及算力紧缺预计将持续 5 到 10 年。 - [04:14](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=254s) Multiverse 运用源自基础量子物理的张量网络算法进行智能剪枝，避免了传统量化压缩容易出现的‘精度断崖’问题。 - [04:14](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=254s) 张量网络压缩技术可稳定实现 50% 到 70% 的模型体积缩减，且精度损失通常控制在 2% 到 3% 以内甚至做到零损失。 - [07:04](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=424s) 在端侧落地案例中，Multiverse 成功将 20B 大模型压缩并塞入 AI PC 预留的 6GB 内存空间中，预计将覆盖数千万台出货设备。 - [07:04](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=424s) 压缩模型使手术机器人、无人机与卫星等边缘设备能够在本地完成实时决策，免去了向云端回传海量数据带来的高昂带宽与推理成本。 - [07:04](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=424s) 部分基础模型在大幅压缩参数与内存占用后，在基准测试中的代码生成等性能反而超越了未压缩的原版，暴露了基础模型中普遍存在的过度参数化和冗余。 - [11:52](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=712s) 商业模式提供模型授权许可（私有部署）与低成本推理即服务 API 两种交付方式，帮助企业遏制飙升至九位数的云端 AI 算力账单。 - [13:42](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=822s) 借助 CompactifAI 智能模型路由器，企业可根据提示词难度和业务目标动态分流请求，往往能在一夜之间将推理成本削减 50% 以上。 - [15:08](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=908s) 推理的未来不是完全依赖英伟达 GPU，通过深度模型压缩，广泛存量的普通 CPU 和新型端侧 NPU 将承载绝大多数日常推理任务。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=4s) 软件行业的 AI 利润危机与持久的算力紧缺 Chris Zaharias（Multiverse Computing 美国区运营负责人）开场介绍了公司背景。Multiverse Computing 是一家拥有 500 名员工的西班牙 AI 企业，近期完成约 5 亿美元融资并成功跻身独角兽行列，其核心业务正是本次演讲的主题：高效与主权 AI 模型。几年前在同一场大会上，团队形成了一个核心论点：AI 训练与算力成本将呈指数级（stratospheric）暴涨，这不仅会带来持续的环境与财务灾难，更将成为各行业的巨大风险。Chris 指出，虽然这一痛点率先在软件行业（ISV）爆发，但未来 5 到 10 年将蔓延至所有产业——让 AI 推理变得极其高效已成为各行各业的生存刚需。 ```mermaid flowchart TD accTitle: AI推理成本对软件企业利润的挤压逻辑 subgraph Squeeze["利润挤压与估值危机"] A["引入 AI 推理功能"] --> B["算力成本与 Token 消耗剧增"] B --> C["企业毛利率严重受挫"] C --> D["投资者下调终值预期与股价暴跌"] end subgraph Supply["算力供给结构性矛盾"] E["算力需求 10-100x 爆发"] --> F["杰文斯悖论压倒摩尔定律"] F --> G["GPU/内存供应紧缺持续 5-10 年"] end D -.-> H["亟需模型压缩与端侧推理方案"] G -.-> H ``` 软件行业正在经历三大现实冲击： - **市值暴跌与利润危机**：过去一年顶尖上市软件公司的市值大幅缩水，根本原因在于软件平台叠加的推理越多，毛利率受侵蚀就越严重，投资人对其预期的终值（terminal value）随之暴跌。很多上市公司 CFO 在投资者大会上被反复追问如何应对 AI 侵蚀利润，却拿不出令人信服的方案。 - **持久的硬件供应危机**：外界普遍预测 GPU、CPU 及内存紧缺会在一两年内缓解，但由于推理需求的爆发式增长，这种算力危机将持续 5 到 10 年。AI 推理能力正变得与招聘人才、基础设施搭建、市场营销同等重要。 - **杰文斯悖论压倒摩尔定律**：尽管摩尔定律在压低单个 Token 的生产成本，但“杰文斯悖论（Jevons paradox）”正在生效——推理成本的边际下降反倒激发了更大规模的推理消耗，推理总需求将彻底碾压硬件效率的自然提升。 --- ### [04:14](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=254s) 源自量子物理的张量网络数学与智能剪枝面对行业面临的算力僵局，Multiverse Computing 提出了一套独特的解决方案。公司总部位于西班牙巴斯克地区（Basque Country，被 Chris 戏称为西班牙的“圣迭戈/拉霍亚”），当地政府长期重资支持基础量子物理研究。公司的四位创始人最初开发了一种用于模拟量子多体状态（quantum many-body states，即亚原子粒子高度相互关联的状态）的全新数学工具。随后他们发现，这套数学体系同样极其擅长分析、理解并剥离神经网络中的冗余与未利用部分。借助这套数学算法，公司在过去一年实现了约 15 倍的业务增长。Chris 调侃自己是“法国文学专业出身”，不会深入推导枯燥的数学公式，但他深入浅出地解释了该技术与传统压缩方法（如全量量化 Quantization）的本质不同： - **传统量化压缩**：通常是全局性地降低模型精度或直接砍掉神经元数量，一旦压缩率超过某个阈值就会遭遇严重的“精度断崖（accuracy cliff）”。 - **张量网络智能剪枝**：通过分析大语言模型（LLM）或神经网络中神经元之间的连接性，精准判断哪些权重与参数对模型实际功能没有贡献并予以剔除。由于越大的模型往往过度参数化（overparameterized）越严重，这为无损压缩留出了巨大的操作空间。该压缩流程分为三步：首先接入数学算法分析模型架构与连接权重，明确各部分的重要性；随后运用张量网络压缩技术剥离冗余；最终可稳定实现 50% 到 60% 甚至高达 70% 的模型压缩，而精度损失通常仅在 2% 到 3% 以内，甚至做到零损失。 --- ### [07:04](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=424s) 落地端侧与 AIPC：从边缘部署到万亿 Token 账单救赎在实际应用方面，Chris 分享了为一家头部 PC 制造厂商提供端侧部署方案的案例。该厂商每年出货 5000 万台 PC，计划在 16GB 内存的 AI PC 上运行本地大模型，并提出了极具挑战性的目标：将一个 200 亿参数（20B）的模型塞进仅分配给它的 6GB 内存空间内。经过数月的工程优化，Multiverse 成功达成目标，该 PC 厂商随后直接对 Multiverse 进行了投资，预计未来一年内将有 4000 万至 5000 万台出货的 PC 搭载该压缩模型。模型压缩带来了广泛的边缘落地与降本能力： - **端侧智能与边缘实时决策**：在手术机器人、普通机器人、无人机、飞机及人造卫星上部署压缩模型，在本地完成数据处理与决策过滤，无需将海量原始数据回传云端数据中心，大幅降低了推理与带宽成本。 - **极速响应与吞吐提升**：如同苍蝇的大脑虽小却反应极快，压缩后的 LLM 和计算机视觉模型不仅大幅减少了物料清单（BOM）中的硬件内存成本，还显著降低了延迟、提高了吞吐量，且大幅减少了对 GPU 和 CPU 算力的依赖。 - **性能甚至超越原版**：在即将发布的 GLM 5.2 压缩版本中，参数量减少 50%、内存占用减少 70%，但在 6 项公开代码评测基准测试中，有 3 项的成绩反而超过了未压缩的原版模型。Chris 指出，目前许多基础模型厂商烧着天文数字的资金，根本不在意开销，其商业模式就是鼓励客户尽可能多地消耗 Token，导致当前基础模型中存在大量的冗余与低效。 - **遏制九位数的算力账单**：在最近三个月里，Chris 接触了大量上市公司，其 Claude、OpenAI 计算及代码生成的综合账单在短短时间内从 7 位数（百万美元）飙升至 9 位数（上亿美元）。公司的业务模式覆盖了两种主要合作形态：一是针对开源与公开模型（如 Qwen、GPT-OSS、DeepSeek、计算机视觉的 YOLO 等累计约 150 款模型）进行压缩，方便企业直接部署于边缘设备或私有化本地环境；二是为企业私有模型定制压缩。例如近期为一家拥有超过 100 款专有模型的软件公司完成了全栈压缩（涵盖 Embedding 模型、语音转文字、文字转语音、视觉语言模型 VLM 等），在仅损失 2% 至 3% 精度的前提下将体积减少 50% 到 60%，有效遏制了客户从 5 亿美元迈向 10 亿美元的云端推理账单。 ### [11:52](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=712s) 商业化交付模式与边缘推理的未来趋势在商业落地方面，Multiverse Computing 主要向企业客户提供两种灵活的合作与交付模式。第一种是**模型授权许可（Model Licensing）**，企业可以直接获得压缩后模型的完整授权，并根据自身的数据合规与架构需求，将其自由部署在私有云、本地数据中心或任何指定环境中；第二种则是**推理即服务 API（Inference-as-a-Service API）**，客户既可以调用原始未压缩的大模型，也可以直接接入经过极限压缩的优化模型。由于压缩模型在 AWS 等云基础设施上消耗的计算资源和底层成本大幅降低，平台能够以远低于常规水平的定价让利给客户，从而成为企业击穿“推理算力危机（Inference Crunch）”的高性价比方案。除了传统的云端 API，推理算力的未来重心正在不可逆转地从中心化云端转移到端侧设备。在接下来的数年内，无论是员工手头的笔记本电脑、家中的电视机顶盒，还是随身携带的智能手机，都将成为承载生成式 AI 推理的核心阵地。对于内部启动了大量生成式 AI 项目、亟需让推理能力更贴近终端设备的企业，模型压缩是实现端侧本地运行的唯一路径。此外，Multiverse Computing 正在与一家以边缘代理服务器与边缘缓存起家的全球知名网络基础设施巨头展开深度合作，将压缩优化后的模型直接内置到其遍布全球的边缘推理云中，推动 AI 推理全面下沉至网络最边缘。 ```mermaid flowchart TD accTitle: Multiverse Computing 交付与推理架构 subgraph Delivery["交付模式"] A["模型授权许可<br/>(自主部署)"] B["推理即服务 API<br/>(低成本云端)"] end subgraph Routing["智能调度"] C["CompactifAI 智能路由器"] end subgraph Hardware["多样化硬件承载"] D["公有云 / 私有云 GPU"] E["全球边缘网络 / CDN 节点"] F["端侧设备 (CPU / NPU)<br/>笔记本 / 手机 / 盒子"] end Delivery --> C C -->|简单/高频请求| F C -->|边缘加速请求| E C -->|超复杂大任务| D ``` --- ### [13:42](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=822s) 智能模型路由：按需分发以立竿见影削减成本在算力紧缺的时代，实现高效推理的另一个关键环节是**智能模型路由（Intelligent Inference Routing）**。虽然行业里可能已有上百家初创公司在推销类似的概念，但智能调度的核心逻辑无可辩驳：现实业务中没有任何一个模型能够或者应该通吃所有任务。无论是参数庞大的旗舰模型还是 OpenAI 的全功能大模型，都不必被大材小用地部署在每一个简单的交互场景中；每种模型在效率、成本和生成质量的光谱上，都有其最适合的运作平衡点。为此，Multiverse Computing 推出了集成自家压缩模型与第三方模型的智能路由器（CompactifAI Router）。该系统能够根据用户输入的提示词内容、用户身份画像以及具体的业务终极目标，自动将请求分流至最具成本效益的模型进行处理。对于正在推进生成式 AI 战略的企业而言，不论是自主研发还是采购第三方方案，智能路由都是必不可少的基础设施组件，企业仅凭一套智能分发机制，往往就能“在一夜之间将推理成本削减一半以上”。 - **模型匹配原则**：依据任务复杂度动态分配，避免“用大炮打蚊子”造成云端算力浪费。 - **降本幅度**：通过智能分发路由，企业在不牺牲输出质量的前提下，推理账单可直降 50% 以上。 - **策略集成**：支持同时接入自主微调模型、Multiverse 压缩模型以及主流商业闭源 API。 --- ### [15:08](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=908s) 硬件范式转移：打破英伟达垄断，重回 CPU 与 NPU 时代关于算力基础设施的演进路线，演讲者提出了一个颠覆主流认知的核心观点："推理的未来不是英伟达 GPU，而是大家已经拥有、并且在全球科技市场中占据绝对主导地位的普通 CPU。" 尽管当前市场狂热追逐英伟达的高端 GPU 算力卡，但真正支撑起全球计算底座的依然是存量巨大的通用处理器。 Multiverse Computing 已经与英特尔（Intel）、高通（Qualcomm）以及 AMD 展开了深入的技术合作与验证。测试表明，只要通过高水准的算法级压缩，就能把业务所需的关键模型顺利塞进“老式的通用 CPU”以及新型端侧 NPU 中平稳运行。在未来的几年里，绝大多数企业级和消费级的日常推理任务，都将摆脱对英伟达顶配 GPU 的绝对依赖，在更普及、更廉价的算力硬件上高效运转。在演讲尾声，发言人邀请现场参会者前往 Multiverse Computing 展位，深入探讨具体业务场景并观看真实压缩模型在端侧与 CPU 上的运行演示。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=4s) AI 训练与算力成本将呈指数级暴涨，这不仅会带来持续的环境与财务灾难，更将成为各行业的巨大风险。 — Chris Zaharias > [00:04](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=4s) 让 AI 推理变得极其高效已成为各行各业的生存刚需。 — Chris Zaharias > [00:04](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=4s) 外界普遍预测 GPU、CPU 及内存紧缺会在一两年内缓解，但由于推理需求的爆发式增长，这种算力危机将持续 5 到 10 年。 — Chris Zaharias > [00:04](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=4s) 推理成本的边际下降反倒激发了更大规模的推理消耗，推理总需求将彻底碾压硬件效率的自然提升。 — Chris Zaharias > [07:04](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=424s) 目前许多基础模型厂商烧着天文数字的资金，根本不在意开销，其商业模式就是鼓励客户尽可能多地消耗 Token，导致当前基础模型中存在大量的冗余与低效。 — Chris Zaharias > [11:52](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=712s) 模型压缩是实现端侧本地运行的唯一路径。 — Chris Zaharias > [13:42](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=822s) 现实业务中没有任何一个模型能够或者应该通吃所有任务。 — Chris Zaharias > [13:42](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=822s) 企业仅凭一套智能分发机制，往往就能在一夜之间将推理成本削减一半以上。 — Chris Zaharias > [15:08](https://www.youtube.com/watch?v=h-e2XJUgZ3o&t=908s) 推理的未来不是英伟达 GPU，而是大家已经拥有、并且在全球科技市场中占据绝对主导地位的普通 CPU。 — Chris Zaharias ## Mentioned Chris Zaharias, Multiverse Computing, CompactifAI, OpenAI, Claude, DeepSeek, Qwen, GPT-OSS, YOLO, GLM 5.2, AWS, Intel, Qualcomm, AMD, NVIDIA ## Video description Chris Zaharias - VP U.S. Operations, Multiverse Computing The biggest threat to enterprise AI ROI isn't model quality it's infrastructure economics. With GPU capacity sold out, hyperscalers rationing access, and agentic workflows multiplying compute consumption 10–100x, companies are hitting a wall. The answer isn't waiting for new chips it's sovereign, efficient AI: compressed models that run where you need them, at the scale you can afford. Multiverse Computing will show how efficient and optimized AI models that reduce their inference costs by 50-80%, giving enterprises, ISVs, and governments a concrete path to AI independence and cost control. [Source](https://www.youtube.com/watch?v=h-e2XJUgZ3o)

Source quotes

[0:04] Chris Zaharias: AI 训练与算力成本将呈指数级暴涨，这不仅会带来持续的环境与财务灾难，更将成为各行业的巨大风险。 [0:04] Chris Zaharias: 让 AI 推理变得极其高效已成为各行各业的生存刚需。 [0:04] Chris Zaharias: 外界普遍预测 GPU、CPU 及内存紧缺会在一两年内缓解，但由于推理需求的爆发式增长，这种算力危机将持续 5 到 10 年。 [0:04] Chris Zaharias: 推理成本的边际下降反倒激发了更大规模的推理消耗，推理总需求将彻底碾压硬件效率的自然提升。 [7:04] Chris Zaharias: 目前许多基础模型厂商烧着天文数字的资金，根本不在意开销，其商业模式就是鼓励客户尽可能多地消耗 Token，导致当前基础模型中存在大量的冗余与低效。 [11:52] Chris Zaharias: 模型压缩是实现端侧本地运行的唯一路径。 [13:42] Chris Zaharias: 现实业务中没有任何一个模型能够或者应该通吃所有任务。 [13:42] Chris Zaharias: 企业仅凭一套智能分发机制，往往就能在一夜之间将推理成本削减一半以上。 [15:08] Chris Zaharias: 推理的未来不是英伟达 GPU，而是大家已经拥有、并且在全球科技市场中占据绝对主导地位的普通 CPU。

Source quotes

[0:00] Chris Zaharias: The biggest threat to enterprise AI ROI isn't model quality it's infrastructure economics. [0:00] Chris Zaharias: With GPU capacity sold out, hyperscalers rationing access, and agentic workflows multiplying compute consumption 10–100x, companies are hitting a wall. [0:00] Chris Zaharias: The answer isn't waiting for new chips it's sovereign, efficient AI: compressed models that run where you need them, at the scale you can afford. [4:00] Chris Zaharias: Over the past year, market capitalizations for top publicly traded software companies have plummeted because layering AI inference onto existing platforms rapidly shrinks gross margins and destroys terminal enterprise value. [4:00] Chris Zaharias: Contrary to optimistic forecasts that memory, GPU, and CPU shortages will resolve within a year or two, Zaharias argues the supply crunch will persist for 5 to 10 years due to exponential inference growth. [4:00] Chris Zaharias: While technological advancements continue to lower token costs, Jevons' paradox will override these gains—as inference becomes cheaper and more accessible, aggregate demand across enterprises will massively outpace hardware efficiency gains. [4:00] Chris Zaharias: By analyzing the complex connectivity between individual neurons, the algorithm identifies which connections genuinely contribute to the network's output and strips away useless weights. [4:00] Chris Zaharias: Reliably reduces model parameter count by 50% to 60% (and up to 95% in extreme cases) with a ~70% reduction in memory footprint and minimal accuracy loss (typically 2% to 3%). [8:00] Chris Zaharias: In a recent project for an enterprise software client managing over 100 proprietary models, Multiverse compressed their embedding, speech-to-text, text-to-speech, and vision-language models by 50% to 60%. [72:00] Chris Zaharias: Implementing an automated query router can instantly cut enterprise inference expenses by over 50%. [72:00] Chris Zaharias: Zaharias predicts that the majority of future AI inference will shift away from high-end Nvidia GPUs toward existing enterprise CPU architectures, NPUs, and hardware from Intel, Qualcomm, and AMD.

Original transcript

[0:04] All [0:04] right. So, I'm going to try to get through in Mary Mer style a bunch of slides, but bring context to [0:09] the whole thing. Um, by way of introductions, my name is Chris Saharius. I run US operations for [0:15] a 500 person Spanish AI company that just recently uh raised about 500 million, got to unicorn status [0:23] on the backs of growth around the topic that I'm going to be presenting today. microphone. Okay, [0:30] so that's the company. Um we're working against one big thesis that we actually learned about at [0:37] this conference several years ago, which is that um training costs, compute costs are going to go [0:43] stratospheric. that it's going to be a sustained uh environmental and financial calamity as well as [0:50] a huge risk factor um initially for one industry that I'm going to talk about but eventually for [0:56] all industries because making uh inference efficient is going to be uh as important in [1:02] all industries uh as as anything ever has been. So I'm going to talk about this in the context of the [1:09] software space uh which is something we've we've studied a lot. We work with a lot of companies, [1:14] a lot of ISVS, and I'm realizing I can't see this, so I'm going to go over here so I can [1:19] see it. Um, so there's three things happening in the software space. Um, market capitalizations of [1:25] top publicly traded software companies have gone through the floor in the last year for precisely [1:31] one reason, which is the more software companies layer inference onto their their platform, [1:38] the more their margins erode and the lower their terminal value, which is the the expected [1:42] value investors have for the company. Um, that's creating a margin crisis. that's going to happen [1:48] in every industry over the next 5 to 10 years, not just software. Uh the second is there's a a huge [1:54] supply crunch which you all know about. A lot of people are are suggesting that [1:59] the memory and GPU and CPU availability crisis should ease in the next year or [2:05] two. Uh we're fairly convinced that it's going to be durable, meaning it's going to last 5, 7, [2:10] 10 years because inference is going to grow so much. So what sounds like a [2:16] you know a theoretical topic AI inference is going to become something that is uh as important as the [2:22] ability to hire employees the as important the ability to have infrastructure uh do marketing. [2:29] Um in this next slide this just sort of illustrates the margin squeeze that AI [2:35] augmented software companies are having to face. So if you go to institutional investor [2:39] conferences for tech and there's a bunch of CFOs of publicly traded software companies, [2:44] a lot of them spend their time answering questions about how are they going to address the fact that [2:49] their great AI solutions that everyone's using are crushing their margins. They don't have a [2:55] they they have an answer, but it's not a real credible one, which is their stocks, [2:58] which is why their stocks are still suffering. Um, we talked about the GPU supply issue [3:04] and and most importantly, I think there's where where while there's Moore's law, which is driving [3:10] down token cost, I think we've all seen that. Um, we think there's every reason to believe that [3:17] uh what's called Jeban's paradox is going to overtake Moore's law. Meaning the amount of [3:22] inference every everyone in every company is is is doing and how much that's going to grow [3:27] is going to absolutely dwarf uh the efficiencies that we can get in in inference. And so this is a [3:35] sustained issue that requires a solution. It can have one of three ends. Uh, one is is ugly, [3:42] which is there's not enough inference. Everyone's every industry's margins suffer or people get more [3:48] efficient. Um, this if you're if you like trading stocks, if you're if you're already gambling here [3:53] in Vegas, this is a a stack ranking of the uh margin pressure um exposure of publicly traded [4:00] software companies uh based on our own internal analysis. So you can see if you look at these [4:05] companies stocks how they've been uh viewed by investors since um since the age of inference. [4:14] We have a solution. It comes from a strange place. Uh we're headquartered in the Basque region in [4:19] Spain. Uh which is kind of like Spain's version of San Diego or La Hoya. Uh the government there has [4:26] always put a lot of money into fundamental quantum physics research and our forefounders [4:31] came up with a new type of math. uh originally built to model quantum manybody states which are [4:37] highly intercorrelated states of subatomic particles. But turns out that math is also [4:42] good at profiling understanding and removing all of the redundancies or the unused portions of [4:50] neural networks uh computer vision models, LLMs, vision language models. And um we built our [4:57] company up around that. We've grown about 15x in the last year. Been able to raise a lot of money. [5:02] growth is good and it's all around this math. And so what I'm going to spend the rest of the the the [5:08] time here talking to you about is what compression and efficiency can mean for your AI initiatives. [5:15] So first of all, I'm not going to try to explain the math. I'm a French lit major so it would [5:19] fail miserably. But I will what I will say is our math is basically uh very different than [5:25] traditional compression methods like quantization in that uh rather than just sort of cutting [5:31] precision across the board or cutting the number of neurons we do what what you could call sort of [5:37] an intelligent pruning which is to say we look at the connectivity between neurons in an LLM or a [5:44] neural network understand which ones do and don't have importance to the functioning of the model [5:50] and then remove um remove weights or parameters that's basically serve no purpose. Uh in this [5:57] world of ever larger models, the larger a model is generally speaking, the more wildly [6:04] overparameterized it is. So it turns out there's a lot of space to make these things smaller without [6:09] sacrificing accuracy. Um we do that through a a sort of three-step approach. just basically stick [6:16] our our math on the model to understand it, how it works, what matters, what doesn't. Uh then use [6:22] tensor network compression to to remove and make the model much much smaller. Uh we can reliably [6:29] compress a model by at least 50 to 60%. Um the big distinction for those of you who've done work with [6:35] quantization and other forms of compression is whereas those methods tend to have a cliff [6:41] an accuracy cliff that that falls off if you go beyond a certain modest level of compression [6:48] we can get to 50 60 70% compression without losing any accuracy or losing as little as 2 to 3%. And [6:56] that's the difference between our method and all other methods. Um we are for example helping a um [7:04] a PC manufacturer that needs an ondevice strategy fit models that they couldn't possibly fit with [7:10] the memory they can they can secure for their 50 million PCs they ship each year. [7:16] They gave us a what sounded like a crazy target which is fitting a 20 billion parameter model in [7:24] six gigs out of a 16 gig AIPC. And uh over the course of a few months we were able to [7:30] do that work. Um and the the results uh speak for themselves. That company's invested in us [7:36] and and you're going to see our models ship with about 40 50 million PCs in the next year. So um [7:43] what can you do with compression? Well, it makes things really efficient. You can place models in [7:48] robots. You can do uh robotic surgery. You can do you can have drones, planes, or satellites [7:55] that rather than having to ship huge amounts of data down to a data center can do processing on [8:00] the device make decisions on what data needs to be shuttled to data centers and dramatically reduce [8:06] overall costs of inference. Um you can also make models faster. So, one of the great things about [8:12] um a FL's brain versus our brains is that it reacts much faster. We do the same thing with [8:18] LLM's computer vision models, so that in addition to taking up a lot less memory, so you don't have [8:24] a a destroyed bill of materials for the product you ship, you can actually get much lower latency [8:30] and you can get higher throughput all while needing less GPUs and CPUs. Um, [8:39] this is what things typically look like. Um, as I said before, I think, um, with our method, [8:45] you can you can very reliably count on at least 50 to 60% compression. We've done as much as 95, [8:51] but I think for most enterprise, uh, generative AI projects, a reasonable goal is 50 to 60%, perhaps [8:59] a bit more. It's about a 70% reduction in memory and um, very little loss. One of the cool things [9:05] about our evolution once we've raised all this money and put it to work with great engineers, [9:11] is we're starting to increasingly come out with models like the compressed version of GLM 5.2, 2, [9:18] which will be out in a couple days, that not only is 50% fewer parameters and 70% smaller in memory, [9:26] but on three of the six coding benchmarks, public coding benchmarks that we measured it against, [9:31] it's actually better than the uncompressed GLM 5.2. So, um to say that there are a [9:37] lot of inefficiencies and opportunities for improvement in foundation models is [9:44] um right now um an understatement. There's a lot of room because they're what are they doing? [9:49] They've taken in ungodly amounts of money. They're not caring at all about what they spend and they [9:55] want you to consume as many tokens as possible. Uh we talked to companies especially in the last [10:01] three months. I've been doing this for two years. In the last three months, the number of companies, [10:06] publicly traded companies I've talked to whose whose aggregate um claude or [10:13] uh open AI compute and coding bill has gone from seven figures to nine figures. Like I've lost [10:19] track of how many of those conversations we've had. So I think we're at a place now where the [10:22] industry is going to get a lot more real and and we look forward to it. Um, our business model is [10:29] compressing any model that we can get our hands on the weights for that matters to customers. So, we [10:33] have our own models we've built. We've compressed everything from Quen to GPTOSS to DeepSeek, [10:41] um, YOLO for computer vision, probably 150 models in total. So if you have a model that you're using [10:49] um we can we either have compressed it or we can compress it so that you can fit it on a device on [10:55] edge in your own onrem environment. Uh we can also take proprietary models. We did a huge project [11:00] for a software company recently that has over a 100 models and they needed to push more of the [11:07] inference from cloud to devices. Uh the result of that was that we we compressed embedding models uh [11:14] speech to text text to speech uh vision language models so that the the overall size of the models [11:21] was 50 to 60% less without anything more than I'd say 2 to 3% precision drop which is going to let [11:29] them take a a cloud inference build that is going from 500 million to a billion dollars this year [11:35] and incrementally reduce that going forward so they can start to answer institutional [11:41] investors who've said they're they're screwed. So that's the the type of thing we work on. Um [11:47] I talked about some of the use cases and how how you can work with us. We have two ways of helping [11:52] companies. One, we license our models. So you you basically take a license of it and you deploy it [11:59] wherever you want to. The other method is we have our own inference as a service API where [12:05] you can use our compressed models. You can also use the uncompressed versions, but the [12:09] benefit of the compressed models is because they cost us a lot less in terms of AWS infrastructure, [12:17] uh we can charge a lot less. So, it's a really good option if you're if you're trying to address [12:21] that uh that inference crunch. We're especially active and I think that one of the great areas [12:30] of the future for inference is it's not going to be in the cloud. We're we're all going to be [12:36] in the next several years getting to a place where you can do inference on your device, [12:43] on your laptop, on your set top box at home, on your phone, that's where things are going [12:49] and I think that's the primary solution. So if uh as a employee at a company that has you know [12:56] myriad uh generative AI initiatives you have edge and you have devices you need to get the inference [13:02] closer to that's something we can help you with. Um we're also partnering with a company everybody [13:08] knows which is a really interesting story in that you know proxy servers they were the first [13:15] ones to cache content on the edge. They got huge because of it. they now have an inference cloud [13:22] um around the world and you can start to think about pushing inference to the edge [13:28] working with them and then you'll find our models uh inside their system. So that's an interesting [13:33] approach as we move progressively from uh from the cloud to to the edge. [13:42] Another piece and and we are by no means unique in this. I think there's it's kind of scary. There's [13:46] probably a hundred companies telling you they have what I'm about to tell you we have. So I'll I'll [13:51] say it. So you know part of efficient inference in the age of the inference crunch is routing [13:57] inference where it needs to. You know fable five fable is not necessary for everything. Open [14:03] open AI is not necessary for everything. There are all sorts of models that each have their [14:08] appropriate use and their appropriate operating point on that spectrum between [14:13] efficiency, cost and uh generation. So we've built um basically uh what we call the capacify router [14:21] which does that work. It's integrated with our models as well as other models and lets you make [14:26] intelligent decisions. I think most of you pro who work on Genai are probably already have something [14:33] or are thinking that you need something and you're either going to build it or get it from someone. [14:38] And whether you get it from us or someone else, I think we we certainly feel like that's a [14:43] critical component for any company's uh AI strategy is to have intelligent routing because [14:49] you can you can typically cut over half the costs of your inference overnight just by [14:55] making sure you have an intelligent process that determines automatically where to route [15:00] queries based on the content of the query and the and the user and the uh the end goal. [15:08] All right. Um, this is a one one point that I don't want to have lost is is we think the future [15:15] of inference is not Nvidia GPUs. It's not Nvidia GPUs. It is the CPUs that you all have already [15:25] and that dominate the global technology market. Uh we've done a lot of work with [15:31] Intel especially, but also um with Qualcomm and with AMD to show that with proper levels [15:38] of compression, you can get the right model that you need for your use case to run [15:43] on good oldfashioned CPUs. You can get them to run on NPUs. And you know the the conclusion is [15:50] I think we're all going to be surprised at the overwhelming majority of inference that's done on [15:57] something other than an the latest and greatest from Nvidia uh in the years to come. [16:05] That is everything uh except the fact that we have an incredible booth [16:09] that we would love for you to stop by and visit. So you we can talk more about your use case, [16:14] show you some examples of our technology in action. And uh with that, I thank you very much.