#AI Infrastructure

Firebird opened its Armenia AI factory, but the 70,000-GPU, 300 MW vision still depends on power, financing and customers. Here is what is actually live today.

Microsoft's expanded multibillion-dollar Mistral pact reverses their 2024 roles — Microsoft now buys European GPU capacity and ships Mistral's models to regulated firms.

NAVER, NVIDIA and Brookfield are tripling Korea's AI factory to 200MW. The real story isn't the megawatts — it's the ~$10 billion capital stack behind sovereign AI.

Alphabet posted record Q2 2026 net income and its first negative free cash flow together. Why $44.9B of AI capex — not the profit — is the real story.

Upper90 will lend General Compute up to $400 million secured by SambaNova inference chips, not Nvidia GPUs. Why inference silicon as loan collateral is an untested bet.

Fireworks AI raised $1.505B at a $17.5B valuation. The number that matters is 95%: nearly all its tokens come from customer-specialized models, not raw open weights.

Google's Ironwood TPU has been generally available since April, yet there's still no public per-hour price. What that silence reveals about AI inference costs.

AMD and Cerebras will split AI inference across Helios racks and the Wafer-Scale Engine. The 5x per-watt claim is modelled against Cerebras-only hardware.

CoreWeave measured 10× more DeepSeek-R1 tokens per megawatt on Vera Rubin NVL72, but missing methodology limits what NVDA investors can conclude today.

Xinference 3.0 turns database-backed authentication on by default and adds an air-gapped deployment path. Here is what operators must test before upgrading.

GPT-5.6 is on Amazon Bedrock, but pricing, context limits, caching and contract terms determine whether AWS commitments create real enterprise savings.

JAX 0.11.0 adds programmable rematerialization for AI training, while a higher dependency floor and changed empty-array behavior raise upgrade risk for teams.

Grok 4.3 reaches Bedrock through Mantle, but runtime incompatibility, regional limits, retention, and retry behavior make enterprise portability conditional.

Thinking Machines Lab's Inkling publishes open weights, but its 600 GB VRAM floor, limited fine-tuning contexts and launch-day gaps complicate ownership.

Memory and chip-equipment stocks crashed on AI glut fears while the Dow hit a record. The new supply lands in the 2030s — here is the math the selloff skipped.

AMD stock dropped 6.9% on July 1, erasing $65 billion, after Meta's cloud plan raised questions about the 6GW MI450 deal that powered AMD's 2026 breakout.

Nvidia stock slipped 1.3% while Meta's cloud plan crushed AI renters and Palantir soared on NVIDIA's sovereign-AI pact. What the two signals mean for NVDA.

Palantir stock added about $22 billion on July 1 after an NVIDIA sovereign-AI pact and an Army data-layer win. The disclosed dollars tell a different story.

Nebius lost about $12 billion in market value on July 1 after Bloomberg reported Meta plans to sell AI compute. Inside the $27B contract that repriced.

Vertiv's AI data-center cooling business has a $15B backlog and rising guidance — but VRT trades near $128B and ~48x forward earnings. The bull and bear case.

Ciena stock shows how AI infrastructure is shifting from chips to bandwidth, with record revenue, a bigger backlog, HyperRail demand, and valuation risk.

Onto Innovation is an AI chip infrastructure stock tied to metrology, inspection, HBM packaging, and gate-all-around process control.

Broadcom's post-earnings drop shows AI demand is still real, but investors now want infrastructure bottlenecks, cash flow, and power access.

SanDisk stock is no longer just a NAND pricing trade. TECHi quote data and SanDisk filings show a $41.6B performance-obligation floor, AI data-center demand, and real valuation risk.