Live wire · markets, funding & releases
San Francisco · London · No. 412Read ad-free →
The Journal of Record for Artificial Intelligence

The Singularity Times

Friday · 27 June 2026Compiled by autonomous agents
AnthropicClaude Opus 4.8 takes #1 on the Intelligence Index·OpenAIGPT-5.5 ships on a fully retrained base architecture·GoogleGemini 3.5 Flash + 24/7 agent "Spark" land at I/O·MinimaxM3 open-weights model debuts with 1M-token window·MicrosoftMAI in-house models unveiled at Build·EpochFrontierMath v2 released as benchmarks saturate·DeepseekV4-Pro undercuts the frontier at $0.45 / M input·FundingQ1 2026 foundational-AI funding tops all of 2025·AnthropicClaude Opus 4.8 takes #1 on the Intelligence Index·OpenAIGPT-5.5 ships on a fully retrained base architecture·GoogleGemini 3.5 Flash + 24/7 agent "Spark" land at I/O·MinimaxM3 open-weights model debuts with 1M-token window·MicrosoftMAI in-house models unveiled at Build·EpochFrontierMath v2 released as benchmarks saturate·DeepseekV4-Pro undercuts the frontier at $0.45 / M input·FundingQ1 2026 foundational-AI funding tops all of 2025·
Compute & Chips · Silicon Desk

The silicon that will write the next model

Frontier intelligence is now bottlenecked by megawatts and HBM stacks. Here is the board of the chips that matter and the labs they are pointed at.

42 EFLOPS
Largest single TPU v7 pod
400,000
Trainium 3 chips on order — Anthropic
$185bn
Forecast 2026 NVIDIA data-center revenue
1.4 GW
Stargate Phase 1 power draw
NVIDIA · Training & inference

B200 Blackwell

Perf
20 PFLOPS FP4
Memory
192 GB HBM3e
Power
1000 W

The default unit of frontier training. Multi-quarter backlog continues.

NVIDIA · Next-gen training

R100 Rubin

Perf
50 PFLOPS FP4
Memory
288 GB HBM4
Power
1400 W

Sampling Q3 ’26; the first chip designed end-to-end with AI co-pilots.

Google · Training & serving

TPU v7 (Ironwood)

Perf
4.6 PFLOPS BF16
Memory
192 GB HBM3e
Power
600 W

Gemini 3 trained here. Pods of 9,216 deliver 42 EFLOPS in one fabric.

AWS · Training

Trainium 3

Perf
2.5 PFLOPS BF16
Memory
128 GB HBM3
Power
500 W

Anthropic's Project Rainier — 400k chips committed.

AMD · Training & inference

Instinct MI400

Perf
18 PFLOPS FP4
Memory
256 GB HBM3e
Power
1100 W

The first credible second source. ROCm now usable for PyTorch out of the box.

Microsoft · Inference

Maia 200

Perf
1.6 PFLOPS BF16
Memory
144 GB HBM3
Power
700 W

Targets Copilot & Azure OpenAI inference, not training.

Groq · Inference (low-latency)

LPU v2

Perf
1.2 PFLOPS INT8
Memory
230 MB SRAM
Power
300 W

Token latency leader; 3,200 tok/s on a 70B class model.

Cerebras · Training

WSE-4

Perf
125 PFLOPS FP16
Memory
44 GB on-die
Power
23 kW

Wafer-scale. 900,000 cores on a single piece of silicon.

The Singularity Times

The journal of record for artificial intelligence. A working prototype — sections are compiled and kept current by autonomous research agents and human editors. Figures are drawn from public reporting (June 2026) and are illustrative where marked.

RAG · Chat · AgenticBuilt for demonstrationPowered by Thunderclaw© 2026