How China's AI Community Is Reshaping the Open Source Computing Stack
China's open source AI models now handle 45% of major routing traffic, as Alibaba Cloud and Cambricon join PyTorch Foundation to reshape global AI.
72 articles
China's open source AI models now handle 45% of major routing traffic, as Alibaba Cloud and Cambricon join PyTorch Foundation to reshape global AI.
Over 80 companies are uniting to standardize physical AI robots, targeting a $200 billion compute opportunity as machines move from labs to the real world.
Equinix, NVIDIA, and Together AI are building a distributed AI inference platform spanning 280 data centers to help enterprises scale AI into production.
CrowdStrike's custom SafeMind AI models beat frontier models by 29% in detection and cut remediation costs by 99%, built on 15 years of security data.
Together AI and Saudi Arabia's HUMAIN are building a 250-megawatt data center projected to generate over $5 billion in AI revenue in year one.
Fashion brands are rethinking costly OpenAI and Anthropic contracts as cheaper open-source models and inference platforms offer a credible alternative.
Nvidia's reported $12.9 billion Hugging Face deal could give the chipmaker control over how millions of developers build and deploy AI models.
Self-hosting AI cuts per-token costs and keeps sensitive data in-house, driving enterprises to move beyond managed APIs from OpenAI and Anthropic.
Kubernetes is becoming the OS for AI factories, letting teams share GPUs safely using DRA, HAMi, and vCluster as AWS and NVIDIA plan 2 million more GPUs.
Together AI topped the fastest-growing startup list by helping startups cut AI costs 80% using open-source models instead of pricey proprietary APIs.
AI inference now needs its own storage layer, and early tests show 90% GPU savings and 20x faster response times using multi-tier SSD caching.
AI safety guardrails are blocking defenders mid-incident while attackers use unconstrained models freely, creating a dangerous asymmetry security teams.
AI sovereignty now demands control over the full infrastructure stack, not just models, as Mistral plans 1 gigawatt of European compute capacity by 2030.
Canva slashed its growth forecast by 10% due to AI model costs, then rebuilt its stack in-house, cutting per-task expenses by 90%.
Nvidia boosted an AI agent's benchmark score from 30% to 100% by improving its harness, proving scaffolding matters more than the model itself.
Harvey's legal AI model Tenet costs a quarter of OpenAI's price, runs on China's Kimi K3, and could reshape how law firms buy AI.
Open-source AI models give every company the same engine; Bridgewater's expert-labeled data cut errors 30% and cost 14 times less than frontier models.
The AI inference market has split into five competing platforms, and picking the wrong one could cost your team far more than the cheapest token price.
AI inference is becoming cloud infrastructure's biggest revenue driver, with CoreWeave's managed inference ARR surging past $100 million in just months.
L&T and Together AI are building India's largest AI factory, a 10,000-GPU cluster worth up to $1.8 billion, to make India a global AI hub.
Together AI now hosts 200 open-source models priced from $0.05 to $9.00 per million tokens, with Chinese labs reshaping the inference market.
IBM chose Together AI over building its own inference engine, signing a $240M deal to deploy open-source AI infrastructure on IBM Cloud in Q1 2027.
Meta's Apache 2.0 licensed AI model runs locally on a 24GB GPU at 233 tokens per second, enabling unrestricted commercial agent deployment without cloud.
Sarvam AI's India-hosted inference platform, not its trillion-parameter model, may be the real key to winning the country's AI infrastructure race.
The break-even point for local AI models vs. cloud APIs has dropped 40%, making self-hosted deployment viable at far lower usage tiers in 2026.
Together AI hosts 200-plus open-weight models with fine-tuning and dedicated GPUs, but teams must choose between its flexibility and Groq's superior.
Cloud providers are now AI's security gatekeepers, as NVIDIA leads a 50-company alliance to make safety a shared infrastructure responsibility.
Red Hat's new Agent Sandbox isolates AI-generated code at the hardware level, solving the three-way trust problem blocking enterprise AI adoption at scale.
AI infrastructure is consolidating fast, as Nscale's Anyscale acquisition and AMD's full-stack push show enterprises now demand end-to-end platforms, not.
Europe will build seven AI Gigafactories backed by €20 billion in private investment to give researchers and startups sovereign computing power for.
Microsoft is competing against OpenAI and Anthropic, companies it invested in, by pushing cheaper homegrown AI models and urging enterprises to avoid.
Open AI model weights don't guarantee enterprise access; Together AI and Fireworks AI fill the cloud gap left by AWS, Azure, and Google on Kimi K3.
Over 95% of tokens served by Fireworks AI come from specialized models, proving the real open-weight opportunity is in customization and serving, not free.
AMD's new Helios AI chips deliver 30% more inference tokens per dollar, with OpenAI, Anthropic, and Meta already committing to large-scale deployments.
No-code AI platforms let non-technical teams build AI apps without coding, and by 2026, 80% of low-code users will work outside formal IT.
Echo's multi-model AI orchestration claims frontier performance at one-third the cost, but mixed benchmarks and opacity over routing decisions fuel.
Legacy code modernization is now the prerequisite for AI adoption, and incremental strategies like the strangler fig pattern can unlock it safely.
Wowza's Video Intelligence Framework turns live camera feeds into real-time AI insights on your own infrastructure, no cloud required, for over one.
Confusing AI-enhanced tools with native AI platforms is costing marketing teams millions; here's how to evaluate AI marketing tools before you invest.
Enterprise analytics teams are ditching legacy platforms for cloud-native, open-source alternatives that cut licensing costs and support modern AI.
AI agents now build complete apps in 90 minutes, autonomously writing code, running tests, and deploying software with minimal human input.
A $400M loan backed by SambaNova inference chips signals capital is fragmenting Nvidia's AI compute dominance, with 16x faster inference promised.
Java developers can now add AI agent interfaces to Spring Boot apps using AG-UI, a standard protocol that eliminates custom WebSocket and event-handling.
40% of AI-generated SQL queries fail silently, inflating revenue figures without warning; here is how teams are fixing the hidden data problem.
Meta's new coding API forces engineering teams to benchmark Muse Spark 1.1 against real workflows, not launch-day claims, before switching providers.
Benchmarks of 90 live AI agent framework runs reveal an 11% token cost gap between top tools, making framework choice a real production decision.
Small AI teams in 2026 are replacing full-time engineers with purpose-built platforms, a survival strategy that matters most when capital is scarce.
SpaceX's $60B Cursor acquisition is reshaping the AI coding tool market, with Cursor losing 15 points of market share within weeks of the deal closing.
91% of mid-market firms use generative AI, but most are ditching off-the-shelf platforms for custom agentic AI development partners to bridge the.
AI agents fail when APIs are poorly designed, and over 40% of agentic AI projects may be canceled by 2027 due to bad infrastructure.