Tamez Labs

Cerebras just built a single computer chip the size of a dinner plate with nearly a trillion transistors, and it runs AI inference up to 15 times faster than Nvidia's best GPUs by eliminating the bottleneck everyone else just accepted. OpenAI liked it so much they signed a 20-billion-dollar multi-year deal and are already using it to power a coding tool, after Cerebras literally took the contract from Nvidia. Now they're filing to go public on Nasdaq, sitting on 24.6 billion dollars in contracted compute, and the chip the entire semiconductor industry spent decades saying was impossible is quietly becoming the backbone of modern AI.

fal just raised 140 million dollars to become the invisible backbone of AI-generated images and videos, processing billions of assets every month for Adobe, Canva, and Shopify. They built an inference engine that makes AI image generation ten times faster than normal, so you get results in under a second instead of waiting around for eight or ten. Their valuation went from 1.5 billion to 4.5 billion in four months because turns out when every app needs to generate media in real time, the company that makes it actually fast enough to work becomes worth a lot of money.

Together AI just dropped a dataset with 258,000 verified coding agent trajectories showing exactly how AI solves real programming problems, and they're running open-source models 2 to 3 times faster than AWS or Google Cloud. The company hit 100 million in revenue and raised 305 million dollars by doing what seemed impossible: making open-source AI infrastructure faster and cheaper than the closed alternatives from OpenAI and Anthropic. Their secret is rewriting the fundamental math operations that power AI models from scratch, led by the same scientist who invented FlashAttention, and now 450,000 developers are using their platform to run over 200 models without managing a single GPU cluster.

Jack Dorsey just fired 4,000 people at Block—40 percent of the company—despite posting 24 percent profit growth and calling results "strong," and Wall Street made the stock jump 20 percent in a single day because he straight-up said AI tools can replace them. He didn't hide behind corporate speak about restructuring or market conditions—he literally told investors that smaller teams using intelligence tools work better, and that most companies are "late" to figuring this out. Block isn't struggling, they're just speed-running the part where profitable tech companies admit humans are now optional if the AI math works.

A startup just raised two billion dollars at seed to let you customize cutting-edge AI models with four lines of Python on your laptop. Thinking Machines Lab's new platform Tinker handles all the nightmare distributed training infrastructure while you write straightforward code, using a technique called LoRA that cuts compute costs by a third compared to traditional fine-tuning. Led by former OpenAI CTO Mira Murati, the company hit a twelve billion dollar valuation out of the gate because they're solving the problem every business hits: general AI models work great in demos but fall apart the second you need them to understand your actual terminology and workflows.

An Ohio police department taught their security cameras to think using AI from a startup you've never heard of, and crime dropped 42 percent while cases got solved 65 percent faster. Reka AI just hit a billion-dollar valuation by doing the opposite of everyone else: building tiny, efficient models that actually work in the real world instead of chasing massive parameter counts. Their 21 billion parameter model beats systems many times its size and can watch video feeds, understand what's happening across multiple cameras, and answer questions like show me everyone in a red jacket who entered this building today.

An AI just beat every human ever on a major software engineering exam in the same two hours, and it's built by a company that cares more about whether you can trust it than how fast it can go. Anthropic's Claude Opus 4.5 just raised thirteen billion dollars while doing something wild: publishing the entire ethical framework that governs how it thinks, making it nearly impossible to trick, and letting it actually do your job across Slack, databases, and design tools. The coding version alone is pulling five hundred million in revenue because developers trust it to build entire features unsupervised—turns out when AI gets powerful enough to matter, safety isn't optional anymore.

The team that built AlphaGo just raised two billion dollars to build AI that actually learns from its mistakes instead of starting fresh every conversation. Reflection AI is combining reinforcement learning with language models to create autonomous coding agents that can write, debug, and fix software without handholding, and they're releasing the whole thing as open-source so anyone can download and own it. It's basically America's answer to China setting the standard for open AI, backed by the guy who wrote the code that beat the world Go champion with that infamous Move 37.

A Brooklyn startup with 60 people just raised 2 billion dollars to build AI models as powerful as GPT-4, then release the actual model weights for free—not just API access, but the stuff OpenAI guards like nuclear codes. Reflection AI is led by the core team behind AlphaGo and Google Gemini, using the same reinforcement learning techniques that beat the world Go champion, except now applied to coding agents that write software that actually works with reliability guarantees. Their first frontier model drops soon, trained on tens of trillions of tokens, and it's backed by Sequoia, NVIDIA, Eric Schmidt, and even the White House AI Czar who said the U.S. needs to win the open-source AI race.

Your brain runs on 20 watts, the same as a dim lightbulb, while the AI systems powering ChatGPT and image generators burn through 50 times more energy doing far less. A company called Unconventional AI just raised 475 million dollars to build chips that run neural networks on physics itself using analog circuits that work like biological neurons, targeting efficiency gains of 1,000 times over current GPUs. The CEO previously sold two AI companies for over a billion dollars combined and walked away from a VP role at Databricks to build this, because AI demand is growing so fast we could literally run out of power to train new models within three to four years.