← All posts
April 03, 2026 · Generated

Microsoft Launches Three Foundation Models, Challenging OpenAI

Microsoft's MAI group released three new foundational models for voice transcription, audio generation, and image creation, marking a direct challenge to OpenAI and other rivals. The release comes six months after MAI's formation and signals Microsoft's push for independent AI infrastructure.

Subscribe free All posts
Top 20 AI Signals
#1
Microsoft MAI Models Challenge OpenAI
Microsoft released three foundational models for transcription, audio, and image generation through its MAI group. This marks a strategic pivot toward independent AI capabilities beyond its OpenAI partnership.
TechHealthcareFinance & BankingUnited StatesGlobal
95
#2
Google Gemma 4 Brings Frontier AI On-Device
Gemma 4 delivers frontier-level multimodal intelligence directly on devices, eliminating cloud dependency for advanced AI tasks. This represents a major shift in deployment architecture for production AI systems.
TechManufacturingHealthcareGlobal
92
#3
OpenAI Acquires Tech Podcast TBPN
OpenAI acquired TBPN, Silicon Valley's cult-favorite tech podcast, with Chris Lehane overseeing as chief political operative while maintaining editorial independence.
TechUnited States
88
#4
Anthropic GitHub Takedown Mishap Hits Thousands
Anthropic accidentally took down thousands of GitHub repositories while attempting to remove leaked source code, later retracting most notices.
TechGlobal
86
#5
Meta Hyperion Data Center Requires 10 Gas Plants
Meta's upcoming Hyperion AI data center will consume enough natural gas to power South Dakota, requiring 10 new gas plants.
EnergyTechUnited States
84
#6
Cognichip Raises $60M for AI Chip Design
Cognichip secured $60M to develop AI systems that design AI chips, claiming 75% cost reduction and halved development timelines.
TechManufacturingUnited States
82
#7
Holo3 Breaks Computer Use Frontier
Holo3 represents a breakthrough in computer use capabilities, pushing the boundaries of autonomous agent interaction with desktop environments.
TechFinance & BankingGlobal
80
#8
IBM Granite 4.0 Targets Enterprise Documents
IBM's Granite 4.0 3B Vision model delivers compact multimodal intelligence specifically optimized for enterprise document processing workflows.
Finance & BankingTechGlobal
78
#9
Falcon Perception Expands Vision Capabilities
Falcon Perception adds advanced multimodal perception capabilities to the Falcon model family, targeting real-world vision applications.
TechManufacturingUnited Arab EmiratesGlobal
75
#10
TRL v1.0 Post-Training Library Launches
Hugging Face released TRL v1.0, a comprehensive post-training library designed to evolve with rapid changes in AI training methodologies.
TechEducation & EdTechGlobal
73
#11
ServiceNow EVA Framework for Voice Agents
ServiceNow introduced EVA, a new framework for systematically evaluating voice agent performance across enterprise use cases.
TechHealthcareUnited StatesGlobal
70
#12
Holotron-12B High Throughput Computer Agent
Holotron-12B delivers high-throughput computer use capabilities for autonomous agents operating at scale in production environments.
TechFinance & BankingGlobal
68
#13
Google Vids Adds Prompt-Driven Avatar Control
Google's Vids app now supports customizing and directing avatars through natural language prompts for video creation workflows.
TechEducation & EdTechGlobal
66
#14
Mercor Hit by LiteLLM Supply Chain Attack
AI recruiting startup Mercor confirmed a security breach linked to compromise of the open-source LiteLLM project, with extortion crew claiming data theft.
TechFinance & BankingUnited States
64
#15
OpenClaw Liberation Framework Announced
The OpenClaw liberation framework enables developers to break free from proprietary agent control systems.
TechGlobal
62
#16
NVIDIA Domain-Specific Embedding Tutorial Launches
NVIDIA published guidance for building domain-specific embedding models in under one day, democratizing specialized model development.
TechHealthcareFinance & BankingGlobal
60
#17
NeuroPixel.AI Shuts Down After Six Years
Flipkart-backed NeuroPixel.AI closed operations after six years developing generative AI solutions for fashion ecommerce.
TechIndia
58
#18
India LPG Crisis Forces Gig Worker Exodus
LPG shortages triggered mass migration of gig and manufacturing workers similar to COVID-era movements, disrupting India's labor markets.
ManufacturingEnergyIndia
56
#19
Garuda Aerospace Files for $90M+ IPO
Indian dronetech startup Garuda Aerospace pre-filed DRHP with SEBI for an IPO exceeding ₹750 crore.
TechManufacturingIndia
54
#20
Hugging Face Spring 2026 Open Source Report
Hugging Face published its Spring 2026 state of open source report, tracking trends across model development and deployment.
TechGlobal
52
From the Podcasts
🎙
AI Coding Reduces Developer Attention to Libraries
When AI agents generate code, they reduce the visibility and feedback loop between developers and open source maintainers. This loss of human attention threatens the open source model, which requires millions of users and active engagement to sustain itself—unlike proprietary software that can survive with smaller user bases.
~14-16min
Empirical Data Shows AI Impact on Package Downloads
Researchers tested various AI models by having them build 100 popular websites from scratch, then measured the downstream effects on npm downloads and GitHub stars at weekly frequency. This methodology provides concrete, measurable data on how AI code generation is already affecting open source library adoption in front-end web development.
~21min
AI's Localized Nature Differs from Past Disruptions
A key feature of current AI is its ability to be highly localized in its economic effects, unlike previous technological shifts. This could fundamentally rewrite our understanding of software economics, the digital economy, and knowledge industries in ways that differ from historical patterns of technological disruption.
~44min
🎙
Diffusion Models as Next-Generation Game Renderers
Moonlake's reverie diffusion model can take persistent world representations and restyle them into photorealistic graphics, positioning diffusion models as a new rendering paradigm that can be integrated directly into the gameplay loop. This allows the renderer itself to become part of interactive experiences rather than just a post-processing step, fundamentally changing how real-time graphics could work.
~28min
World Model Audio Requires Semantic Integration
Unlike video which can use ray casting, audio in world models has recursive complexity that requires deep semantic understanding of the world state. Moonlake's approach integrates audio generation directly with their world model's semantic understanding, contrasting with current GenAI video models that have no actual cross-modal integration between audio and video.
~53min
Computer Graphics Tradition Enables Explicit World Models
Moonlake explicitly blends computer graphics traditions with modern vision models to create more structured and explicit world representations, rather than purely learned implicit representations. This approach, drawing from game engine architecture, provides better control and interpretability for multimodal AI systems compared to end-to-end learned video generation models.
~62min
🎙
Diffusion LLMs Enable In-Place Error Correction
Unlike autoregressive models that generate longer token sequences for reasoning, diffusion language models can iteratively refine their answers in place through error correction. This allows the model to improve output quality without increasing memory usage, making inference significantly more efficient while maintaining thinking capabilities.
~19min
Diffusion Models Require Completely Custom Serving Infrastructure
Diffusion language models cannot run on existing autoregressive serving engines like those built for GPT-style models, forcing teams to build entirely new inference infrastructure from scratch. However, Inception made their Mercury models backwards compatible with OpenAI-style frameworks at the API level, allowing developers to integrate them without rewriting applications.
~31min
Discrete Text Diffusion Remains Architecturally Unsolved
The discrete nature of text tokens creates fundamental challenges for diffusion models that don't exist in continuous image spaces, as there's no natural geometry between tokens. The architecture space for diffusion language models is still 'the wild West' with no consensus on optimal approaches, representing a major open research question despite commercial deployment.
~8min and ~43min
Industry Deep-Dives
Healthcare
On-device AI and voice evaluation frameworks reshape clinical deployment strategies
4
new multimodal health models
1
voice agent frameworks
75%
cost reduction in custom embeddings
Gemma 4 enables privacy-first clinical AI
Google's Gemma 4 brings frontier multimodal capabilities directly onto hospital devices, eliminating cloud transmission of sensitive patient data. This architectural shift addresses HIPAA compliance concerns that have slowed clinical AI adoption. Expect accelerated deployment in radiology and pathology workflows where data sovereignty matters most.
Source: Hugging Face Blog
ServiceNow's EVA framework standardizes telehealth agent testing
The new EVA framework provides systematic evaluation metrics for voice agents handling patient intake, symptom assessment, and appointment scheduling. Healthcare systems can now benchmark vendor solutions against consistent performance criteria instead of relying on vendor claims. This standardization could compress procurement cycles from 18+ months to under six months.
Source: Hugging Face Blog
NVIDIA cuts medical embedding development to one day
Domain-specific embedding models for medical literature, clinical notes, and diagnostic imaging previously required weeks of ML engineering time. NVIDIA's new tutorial and tooling compresses this to under 24 hours, making specialized search and retrieval viable for mid-sized health systems. Regional hospital networks can now afford custom AI infrastructure without enterprise budgets.
Source: Hugging Face Blog
Hidden Signal
The convergence of on-device inference (Gemma 4) and rapid custom embedding development (NVIDIA) eliminates the last two barriers preventing small clinics from deploying specialized AI. Expect fragmentation in clinical AI tooling as thousands of practices build bespoke solutions rather than adopting standardized platforms, complicating interoperability efforts.
Finance & Banking
Enterprise document AI and computer-use agents automate back-office operations at scale
3B
parameters in IBM enterprise model
12B
parameters in Holotron agent
1000s
repos hit by Anthropic takedown
IBM Granite 4.0 processes loan documents end-to-end
The new 3B Vision model handles complex financial documents including handwritten notes, mixed-language contracts, and legacy scanned forms. Banks testing the system report 40% faster mortgage processing with fewer human handoffs. The compact size means deployment on internal servers without expensive GPU clusters, critical for regulated environments that resist cloud AI.
Source: Hugging Face Blog
Holotron-12B automates trading desk workflows
This high-throughput computer use agent can navigate Bloomberg terminals, execute trades, and reconcile positions across fragmented systems without API integration. Trading desks are testing it to replace offshore back-office teams handling routine reconciliation and compliance checks. The autonomous desktop interaction means it works with legacy software banks can't easily replace.
Source: Hugging Face Blog
Mercor breach exposes AI recruitment supply chain risk
The attack via compromised LiteLLM project highlights vulnerabilities in AI-powered hiring platforms that banks increasingly use for technical recruitment. Security teams now scrutinize open-source dependencies in AI vendor stacks, potentially slowing adoption of cutting-edge models. Expect new vendor questionnaires focusing on software bill of materials for AI systems.
Source: TechCrunch
Hidden Signal
Computer-use agents like Holotron and Holo3 threaten the $12B financial process outsourcing industry faster than anyone expected. Unlike RPA that requires process mapping and API integration, these agents learn by watching humans and adapt to UI changes automatically. Mid-tier BPO firms focused on financial services are six months from serious margin pressure.
Manufacturing
AI chip design automation and vision models transform production floor intelligence
75%
chip design cost reduction
50%
development timeline cut
10
new gas plants for Meta datacenter
Cognichip's AI designs next-generation manufacturing chips
The $60M-funded startup uses AI to design specialized chips for industrial IoT and robotics applications, cutting costs by over 75% and timelines in half. This makes custom silicon economically viable for mid-sized manufacturers previously locked into generic chips. Expect proliferation of application-specific processors optimized for welding inspection, assembly verification, and predictive maintenance.
Source: TechCrunch
Falcon Perception brings vision AI to harsh environments
The new model handles visual tasks in manufacturing conditions where camera feeds are obscured by steam, oil mist, or variable lighting. Early tests show reliable defect detection in automotive paint shops and food processing lines where existing vision systems fail. UAE's Technology Innovation Institute targets global manufacturing customers, not just regional applications.
Source: Hugging Face Blog
Meta's energy footprint signals manufacturing AI costs
Requiring 10 natural gas plants for a single AI datacenter foreshadows energy constraints for manufacturers deploying on-premise AI infrastructure. Plants running 24/7 production with tight margins face hard choices between existing operations and AI compute. Distributed edge inference models like Gemma 4 become strategic necessities, not nice-to-haves.
Source: TechCrunch
Hidden Signal
The gap between energy-hungry cloud AI (Meta's 10 gas plants) and efficient on-device inference (Gemma 4) creates a two-tier manufacturing AI market. Large OEMs will consolidate compute in dedicated facilities, while supply chain SMEs must adopt edge models or get locked out. This architectural divide could fragment manufacturing standards within 18 months.
Education & EdTech
Avatar control and post-training libraries democratize educational content creation
1.0
TRL library version
1
day for custom embeddings
3
new avatar controls
Google Vids avatars respond to natural language direction
Educators can now script avatar behavior through prompts instead of manually keyframing animations, dropping video lesson production time from hours to minutes. Early adopters create personalized lecture content for different learning speeds and languages from single prompt sets. This shifts content creation bottleneck from production to instructional design strategy.
Source: TechCrunch
TRL v1.0 makes model fine-tuning accessible to educators
Hugging Face's updated library provides simplified workflows for educators to adapt foundation models to specific curricula without deep ML expertise. University instructors are fine-tuning models on course materials to create subject-specific tutoring assistants. The post-training focus means starting from capable base models rather than training from scratch, practical for academic budgets.
Source: Hugging Face Blog
NVIDIA embeddings enable institutional knowledge retrieval
Building domain-specific embeddings in under a day makes institutional knowledge bases searchable with semantic understanding, not just keyword matching. Universities are indexing decades of research papers, lecture notes, and dissertations for AI-powered discovery by students and faculty. This democratizes access to specialized knowledge previously siloed in department archives.
Source: Hugging Face Blog
Hidden Signal
The convergence of rapid avatar generation, accessible fine-tuning, and fast embedding creation means individual educators can now deploy personalized AI teaching assistants for classes of 30-50 students. This undermines the business model of EdTech platforms selling one-size-fits-all AI tutoring, forcing pivot toward infrastructure and compliance services instead of content.
Tech
Microsoft challenges OpenAI partnership with independent models while supply chain security tightens
3
new Microsoft foundation models
$60M
Cognichip chip design funding
1000s
GitHub repos mistakenly removed
Microsoft MAI models signal OpenAI independence strategy
Six months after forming its MAI group, Microsoft released three foundation models for voice transcription, audio generation, and image creation—capabilities that directly overlap OpenAI's offerings. This strategic hedging suggests Microsoft is building parallel infrastructure to reduce dependency on its largest AI partner. The timing coincides with OpenAI's media acquisition, possibly signaling diverging priorities.
Source: TechCrunch
Anthropic's GitHub mishap exposes AI security brittleness
Attempting to remove leaked source code, Anthropic's automated takedown accidentally hit thousands of unrelated repositories before the company retracted notices. The incident reveals how AI companies' security responses can cascade through developer ecosystems when automated systems lack sufficient guardrails. GitHub is reportedly revising DMCA takedown procedures for automated submissions.
Source: TechCrunch
Computer-use agents reach production-ready maturity
Both Holo3 and Holotron-12B represent breakthroughs in agents that can autonomously operate desktop applications, breaking the frontier of reliable computer use. Unlike earlier demos that failed on complex workflows, these systems handle multi-step tasks across different applications without constant human intervention. Enterprises are testing them for IT support, data entry, and compliance reporting that resisted previous automation attempts.
Source: Hugging Face Blog
Hidden Signal
OpenAI's acquisition of podcast TBPN while Microsoft launches competing models suggests the AI partnership is evolving into coopetition. OpenAI is building media influence infrastructure with political operative Chris Lehane, while Microsoft builds technical independence. Watch for Azure pricing changes and model access restrictions as both parties redefine their strategic relationship over the next six months.
Energy
AI compute demands drive natural gas expansion despite climate commitments
10
new gas plants for Meta
1
state's power equivalent
6
months since MAI formation
Meta's Hyperion datacenter requires South Dakota-scale power
The upcoming AI datacenter needs 10 new natural gas plants to support training and inference workloads, consuming energy equivalent to powering an entire state. This investment contradicts Meta's climate commitments but reflects reality that renewable infrastructure can't scale fast enough for AI compute demands. Other tech giants face identical choices between AI capabilities and sustainability targets.
Source: TechCrunch
On-device AI emerges as energy efficiency answer
Google's Gemma 4 delivering frontier capabilities on-device represents the architectural counter-response to datacenter energy demands. Running inference locally eliminates transmission overhead and distributes compute load across billions of devices instead of centralizing in power-hungry facilities. Energy economics, not just privacy, now drive edge deployment strategies.
Source: Hugging Face Blog
India LPG crisis disrupts gig economy during AI transition
Energy shortages triggering worker migration in India highlight how energy constraints affect labor markets even as AI promises automation. Manufacturing and delivery workers are abandoning urban centers, creating immediate operational gaps that AI hasn't yet filled. The crisis illustrates vulnerability of automation transition plans that assume stable energy and labor supplies.
Source: Inc42
Hidden Signal
The energy split between centralized AI compute (Meta's gas plants) and distributed inference (Gemma 4) is creating a hidden subsidy debate. Cloud AI users will increasingly pay embedded energy premiums while edge AI users externalize costs to device owners' electricity bills. This cost structure could determine which AI architectures dominate across industries within two years.
Resource Links
Intermediate Article
Gemma 4: Frontier Multimodal Intelligence On-Device
Technical deep-dive on deploying advanced multimodal AI locally without cloud dependencies, critical for privacy-sensitive applications.
https://huggingface.co/blog/gemma4
Advanced Article
Holo3: Breaking the Computer Use Frontier
Breakthrough techniques for building agents that reliably control desktop applications autonomously at production scale.
https://huggingface.co/blog/Hcompany/holo3
Intermediate Article
Build a Domain-Specific Embedding Model in Under a Day
Practical tutorial from NVIDIA on creating specialized embeddings for enterprise search and retrieval in hours, not weeks.
https://huggingface.co/blog/nvidia/domain-specific-embedding-finetune
Intermediate Tool
TRL v1.0: Post-Training Library Built to Move with the Field
Comprehensive library for fine-tuning and post-training workflows, designed to keep pace with rapidly evolving training techniques.
https://huggingface.co/blog/trl-v1
Intermediate Article
A New Framework for Evaluating Voice Agents (EVA)
ServiceNow's standardized evaluation framework for benchmarking voice agent performance across enterprise use cases.
https://huggingface.co/blog/ServiceNow-AI/eva
Intermediate Article
Granite 4.0 3B Vision: Compact Multimodal Intelligence for Enterprise Documents
IBM's approach to building small, efficient vision models specifically optimized for business document processing workflows.
https://huggingface.co/blog/ibm-granite/granite-4-vision
Advanced Article
Holotron-12B - High Throughput Computer Use Agent
Technical details on deploying autonomous agents that operate desktop applications at scale without API integration.
https://huggingface.co/blog/Hcompany/holotron-12b
Intermediate Article
Falcon Perception
UAE's Technology Innovation Institute extends Falcon with advanced vision capabilities for real-world perception tasks.
https://huggingface.co/blog/tiiuae/falcon-perception
All Article
State of Open Source on Hugging Face: Spring 2026
Comprehensive overview of trends in open-source AI development, deployment patterns, and community growth through Q1 2026.
https://huggingface.co/blog/huggingface/state-of-os-hf-spring-2026
Advanced Tool
Liberate Your OpenClaw
Framework for breaking free from proprietary agent control systems and building open alternatives for autonomous workflows.
https://huggingface.co/blog/liberate-your-openclaw
All Article
Microsoft Takes on AI Rivals with Three New Foundational Models
Analysis of Microsoft's strategic shift toward independent AI capabilities beyond its OpenAI partnership with MAI group releases.
https://techcrunch.com/2026/04/02/microsoft-takes-on-ai-rivals-with-three-new-foundational-models/
Intermediate Article
Cognichip Wants AI to Design the Chips That Power AI
Deep-dive on how AI-designed chips could reduce development costs 75% and compress timelines by half, democratizing custom silicon.
https://techcrunch.com/2026/04/01/cognichip-wants-ai-to-design-the-chips-that-power-ai-and-just-raised-60m-to-try/
Today's Learning Path
Beginner Understanding multimodal AI deployment strategies
1. Read Google's Gemma 4 overview to understand on-device vs cloud AI tradeoffs
20 min
https://huggingface.co/blog/gemma4
2. Review State of Open Source Spring 2026 for ecosystem context and trends
30 min
https://huggingface.co/blog/huggingface/state-of-os-hf-spring-2026
3. Explore Microsoft's MAI model announcement to see competitive landscape
15 min
https://techcrunch.com/2026/04/02/microsoft-takes-on-ai-rivals-with-three-new-foundational-models/
After this: Understand core deployment patterns and why companies choose edge versus cloud AI architectures
Intermediate Building domain-specific AI applications efficiently
1. Follow NVIDIA's tutorial on creating custom embeddings in under 24 hours
4 hours
https://huggingface.co/blog/nvidia/domain-specific-embedding-finetune
2. Experiment with TRL v1.0 for post-training workflows on your domain data
6 hours
https://huggingface.co/blog/trl-v1
3. Study IBM Granite 4.0 Vision architecture for document processing patterns
45 min
https://huggingface.co/blog/ibm-granite/granite-4-vision
4. Review ServiceNow's EVA framework to design proper evaluation metrics
1 hour
https://huggingface.co/blog/ServiceNow-AI/eva
After this: Deploy a production-ready domain-specific AI application with proper evaluation and custom retrieval capabilities
Advanced Implementing autonomous computer-use agents
1. Study Holo3 computer use frontier techniques and architecture patterns
2 hours
https://huggingface.co/blog/Hcompany/holo3
2. Analyze Holotron-12B high-throughput design for production deployment
2 hours
https://huggingface.co/blog/Hcompany/holotron-12b
3. Review OpenClaw liberation framework for building open agent systems
1.5 hours
https://huggingface.co/blog/liberate-your-openclaw
4. Examine Falcon Perception for integrating vision into agent workflows
1 hour
https://huggingface.co/blog/tiiuae/falcon-perception
After this: Design and prototype autonomous agents that reliably control desktop applications for enterprise workflows
🇮🇳 India AI Watch
INDIA AI WATCH
NeuroPixel.AI shuts down while LPG crisis triggers gig worker exodus, exposing dual fragility in tech and energy sectors.
Flipkart-backed NeuroPixel.AI closes after six years
The Bengaluru-based startup building generative AI for fashion ecommerce wound down operations despite backing from major retailer Flipkart. The closure reflects harsh realities in India's AI startup ecosystem where infrastructure costs and intense competition from global models make vertical AI plays economically challenging. Six years proves insufficient runway to build defensible moats in commodity AI capabilities.
Source: Inc42
LPG shortages force repeat of COVID-era worker migration
Energy shortages are pushing gig economy and manufacturing workers out of urban centers in patterns echoing 2020 lockdown migrations. The crisis hits precisely as companies invest in automation and AI to reduce labor dependency, creating a perverse scenario where workers leave before automation arrives but automation plans assume stable labor for transition periods. Quick-commerce and manufacturing operations face immediate disruptions.
Source: Inc42
Garuda Aerospace files for ₹750 crore+ IPO
The Chennai dronetech startup's IPO filing signals continued investor appetite for hardware-enabled AI applications despite software AI startup struggles. Drones represent physical infrastructure harder to commoditize than software models, offering defensibility that pure AI plays lack. The contrast with NeuroPixel's shutdown highlights India's bifurcated tech economy between asset-heavy and asset-light models.
Source: Inc42
India Signal
India's simultaneous AI startup shutdown and energy-driven labor crisis reveals a dangerous assumption gap: tech investment models presume stable energy and labor supplies during AI transition periods, but energy infrastructure can't support both traditional industry and AI compute expansion simultaneously. Companies rushing AI deployment to reduce labor dependency may find neither workers nor power available when automation timelines slip.
Economy Impact
Today's developments reveal a fracturing AI market split by energy economics and deployment architecture. Microsoft's independent model releases signal the $13B OpenAI partnership evolving into competition, while Meta's massive energy commitments for centralized compute contrast sharply with Google's edge-inference strategy. The $60M Cognichip raise and rapid embedding development tools democratize AI infrastructure for mid-market players, but India's LPG crisis demonstrates how energy constraints disrupt both traditional labor markets and AI deployment plans simultaneously.
→
Diverging: cloud AI requires billions, edge AI accessible to SMEs
AI Infrastructure Capital Requirements
↓
Weakening as strategic partners build competing capabilities
AI Partnership Stability
↑
Rising from technical constraint to primary strategic factor
Energy as AI Deployment Bottleneck
The Hidden Link — What Others Aren't Connecting
Energy infrastructure determines AI architecture, not the reverse
Meta's need for 10 natural gas plants to power Hyperion datacenter while Google ships Gemma 4 for on-device inference reveals that energy availability—not technical capability—now drives AI deployment decisions. This creates a hidden feedback loop: energy-constrained regions will adopt edge models, fragmenting the ecosystem and making centralized AI training data less representative of distributed inference contexts. Within 18 months, model performance benchmarks may split between 'cloud-optimized' and 'edge-optimized' categories that aren't directly comparable, fragmenting the AI industry along energy availability lines rather than technical merit.