Erick Johnson - The Silicon Greenhouse

The Silicon Greenhouse

Sign in Subscribe

Erick Johnson

NVIDIA's H100X: Why 94 GB of Memory Changes Everything

NVIDIA's H100X: Why 94 GB of Memory Changes Everything

NVIDIA is quietly preparing the H100X, a refined Hopper GPU with 94 GB of memory. Here's why that extra 14 GB matters more than raw compute.

Kimi K3 Is So Popular, Moonshot Ran Out of GPUs

Kimi K3 Is So Popular, Moonshot Ran Out of GPUs

Just 48 hours after launching Kimi K3, Moonshot AI suspended new subscriptions as demand overwhelmed their GPU cluster. Here's when signups might resume and what the July 27 open-weight release means for the capacity crunch.

Qwen 3.8: Alibaba's 2.4 Trillion-Parameter Bet on Open Weights

Qwen 3.8: Alibaba's 2.4 Trillion-Parameter Bet on Open Weights

Alibaba announced Qwen 3.8, a 2.4T parameter model promised to be 'second only to Fable 5' — but without a single published benchmark. Here's how it stacks up against Kimi K3, GLM 5.2, and DeepSeek V4 in the increasingly competitive open-weight race.

The U.S. Wants a FINRA for AI: What That Actually Means

The U.S. Wants a FINRA for AI: What That Actually Means

The Trump administration is exploring an industry-funded AI watchdog modeled after Wall Street's FINRA. Here's how it would work and why some AI labs actually support it.

Federal Overreach: How the Trump Administration Is Dismantling State AI Laws

Federal Overreach: How the Trump Administration Is Dismantling State AI Laws

The Trump administration is using financial threats, litigation, and executive orders to dismantle state-level AI regulations. From funding penalties to temporary preemption periods, here's how federal pressure is reshaping AI governance.

GPT-5.6 Gets Federal Approval After 12-Day White House Review

GPT-5.6 Gets Federal Approval After 12-Day White House Review

OpenAI's GPT-5.6 cleared a 12-day federal review involving White House officials, marking a shift from voluntary AI safety commitments to government-managed access for frontier models.

NVIDIA's H200: The Memory Monster Upgrading AI Infrastructure Without Starting Over

NVIDIA's H200: The Memory Monster Upgrading AI Infrastructure Without Starting Over

The H200 packs 141GB of memory and 4.8 TB/s bandwidth into the familiar Hopper architecture—offering a practical upgrade path for data centers not ready to jump to Blackwell.

The AI Gold Rush Hits Wall Street: Inside the Multi-Trillion Dollar IPO Wave

The AI Gold Rush Hits Wall Street: Inside the Multi-Trillion Dollar IPO Wave

As SpaceX and xAI shatter records with a historic $2.1 trillion public debut, Anthropic and OpenAI prepare their own blockbuster IPOs. This market breakdown analyzes how these pure-play AI valuations stack up against tech giants like Alphabet.

HHS Is Deploying ChatGPT to Hunt for $200 Billion in Medicaid Fraud. What Could Go Wrong?

HHS Is Deploying ChatGPT to Hunt for $200 Billion in Medicaid Fraud. What Could Go Wrong?

The HHS is using ChatGPT to scan five years of state Medicaid audits to find up to $200 billion in fraud. But critics warn that AI hallucinations and a lack of human nuance could turn the tool into an automated judge and executioner.

The $2 Trillion Rocket: Inside the SpaceX IPO and Musk’s New Financial Stratosphere

The $2 Trillion Rocket: Inside the SpaceX IPO and Musk’s New Financial Stratosphere

SpaceX just shattered records with a historic $75 billion public debut, pushing its valuation past $2 trillion and making Elon Musk the world's first trillionaire. Here is what is actually inside the company's hidden empire.

VRAM Sweet Spot: Getting the Most Out of a 16GB GPU for Local AI

VRAM Sweet Spot: Getting the Most Out of a 16GB GPU for Local AI

Discover the true limits of a 16GB GPU in local AI. From running quantized 16B language models to mastering Flux image generation and cinematic video generation.

Z.ai Releases GLM-5.2: 1M Context, MIT License, and the End of Compute Bloat

Z.ai Releases GLM-5.2: 1M Context, MIT License, and the End of Compute Bloat

Z.ai has officially released GLM-5.2, featuring a 1-million-token context window and an MIT open-source license. The 753B parameter model introduces architectural tricks like IndexShare to dramatically lower compute overhead.

Venice AI: Image Generation Showdown

Venice AI: Image Generation Showdown

We put 10 Venice AI image models—from free SDXL checkpoints to premium models like Flux 2 Max, GPT Image 2, and Grok—through a brutal 3-prompt stress test. No cherry-picking, single-shot results. See which model actually wins on value.

The Federal Recall of Claude Fable 5: Safety First or Political Payback?

The Federal Recall of Claude Fable 5: Safety First or Political Payback?

Anthropic was forced to abruptly disable its new Fable 5 and Mythos 5 models following a sudden 90-minute federal directive. The unprecedented move has sparked massive debate over government overreach, security vulnerabilities, and the growing geopolitical tech race with China.

OpenAI Faces Massive 42-State Subpoena Just Days After Secret IPO Filing

OpenAI Faces Massive 42-State Subpoena Just Days After Secret IPO Filing

A massive 42-state coalition of attorneys general has hit OpenAI with a sweeping subpoena targeting child safety, data privacy, and deceptive advertising. The sudden legal blitz lands just days after the AI giant filed for its highly anticipated $850 billion IPO.

Beyond the Prompt: Why 2026 Is the Year of the Autonomous AI Agent

Beyond the Prompt: Why 2026 Is the Year of the Autonomous AI Agent

The era of the passive chat prompt is ending as major tech labs ship true autonomous infrastructure. Here is how OpenAI, Anthropic, and Google are turning AI agents into independent coworkers and rewriting the rules of SaaS.

Meta’s AI Pivot Proves That "Efficiency" Is Just Big Tech Code for Firing People

Meta’s AI Pivot Proves That "Efficiency" Is Just Big Tech Code for Firing People

Mark Zuckerberg's recent internal memo admits to mistakes during Meta's massive AI restructuring. As 8,000 employees lose their jobs to fund a $145 billion hardware bill, the true cost of Big Tech's AI efficiency narrative is exposed.

Oracle's Quiet Nerf to the Always Free Tier: What You Need to Do Before June 15th

Oracle's Quiet Nerf to the Always Free Tier: What You Need to Do Before June 15th

Oracle has quietly halved its Always Free ARM tier from 4 CPUs and 24GB RAM to 2 CPUs and 12GB RAM. With changes hitting on June 15th, free instances face shutdowns and PAYG users face sudden charges.

Microsoft and NVIDIA Just Changed the PC Architecture Board Forever

Microsoft and NVIDIA Just Changed the PC Architecture Board Forever

Microsoft and NVIDIA's new RTX Spark partnership brings data-center-class AI processing directly to Windows laptops. By running massive 120B models locally, this architecture fundamentally shifts the power dynamic away from Apple and cloud-only APIs.

Moonshot AI Drops Kimi K2.7-Code: Smarter Software Engineering That Won't Break the Bank

Moonshot AI Drops Kimi K2.7-Code: Smarter Software Engineering That Won't Break the Bank

Moonshot AI has launched Kimi K2.7-Code, an open-weight coding model designed to maximize token efficiency. By slashing thinking tokens by 30% while matching Kimi K2.6 pricing, it offers developers a highly affordable alternative to heavyweights like GLM 5.1 and DeepSeek V4 Pro.

Robinhood Opened the Gates to Agentic AI Trading. Is it Time to Hand Over Your Portfolio?

Robinhood Opened the Gates to Agentic AI Trading. Is it Time to Hand Over Your Portfolio?

Robinhood has launched an MCP-powered Agentic Trading feature, allowing external AI models like Claude and ChatGPT to trade stocks on your behalf. Here is a look at how it works, the safety guardrails, and the early feedback from traders.

Unshackling the Mythos: How Claude Fable 5 Alters the 2026 Frontier AI Landscape

Unshackling the Mythos: How Claude Fable 5 Alters the 2026 Frontier AI Landscape

Anthropic has officially launched Claude Fable 5, the first public model of its ultra-powerful Mythos tier. Here is how it stacks up against GPT-5.5, Gemini 3.1 Pro, and DeepSeek V4 Pro.

Why I Ditched the Heavyweights and Switched My AI Agent to DeepSeek V4

Why I Ditched the Heavyweights and Switched My AI Agent to DeepSeek V4

Running autonomous AI agents can quickly break the bank on mainstream APIs. Switching to the dual-tier DeepSeek V4 architecture cuts daily operation costs to pennies without sacrificing deep reasoning.

MiniMax M3: The 1M Context Beast That Just Reset the Coding Benchmarks

MiniMax M3: The 1M Context Beast That Just Reset the Coding Benchmarks

MiniMax M3 has arrived with a massive 1M context window and benchmark scores that challenge GPT-5.5. We dive into the specs, API pricing, and why this model is a new favorite for AI agents.

The Long Game: Why Raspberry Pi 6 is Staying "Dumb" to Stay Smart

The Long Game: Why Raspberry Pi 6 is Staying "Dumb" to Stay Smart

The Raspberry Pi 6 is officially aimed for a 2028 release, and it’s skipping the NPU trend. Discover why Eben Upton is prioritizing modularity and affordability over the AI hype.