TickerTain
TickerTain
NewsroomShortsPortfolioConvergence
NewsroomShortsPortfolioConvergence
$ANTHROPIC·$MA····$INTC····$BLUE-ORIGIN·$SPCX····$CRWV····$CRM····$MSFT····$NVDA····$ORCL····$CURSOR·$AAPL····$OPENAI·$AMZN····$UBER····$GOOGL····$META····$TSLA····$DATABRICKS·$PERPLEXITY·$LYFT····$NBIS····$TSM····$LITE····$ANDURIL·
$ANTHROPIC·$MA····$INTC····$BLUE-ORIGIN·$SPCX····$CRWV····$CRM····$MSFT····$NVDA····$ORCL····$CURSOR·$AAPL····$OPENAI·$AMZN····$UBER····$GOOGL····$META····$TSLA····$DATABRICKS·$PERPLEXITY·$LYFT····$NBIS····$TSM····$LITE····$ANDURIL·
←

immad akhund

T2 · manager / operator

Immad Akhund is an entrepreneur and author of 'The Last Economy', deriving economics from generative AI math. He leads research on quantization, edge deployment, and AI forecasting, and runs an AI startup.

3 calls·3 names·100% bull·last heard 14 days ago·Peter H. Diamandis
track record

no scored calls yet — needs a stated position or a categorical verdict, with a matured window vs SPY

top calls

highest conviction · one per company
1sthigh conviction
$DEEPSEEKDeepSeek

DeepSeek Flash model slashes KV cache memory 50x (48K to 890) and beats GPT-4/Opus at 20x lower cost

Chinese labs are algorithmically routing around HBM constraints via sparsity and encoder-decoder innovations — DeepSeek's $10M training cost vs $1B+ for Western models proves a 10-15x cost advantage for fast followers using distillation, reshaping the inference economics and fab investment thesis.

Peter H. Diamandis2026-09episode →
2ndmedium conviction
$TCEHYTencent

Akhund: Tencent's Hunyuan team achieves binary quantization for 300B models on consumer hardware

Tencent's WizardLM-derived team broke the 1-bit barrier, enabling 300B parameter models on MacBooks — a milestone for edge AI deployment.

Peter H. Diamandis2026-07episode →
3rdlow conviction
$FIREWORKS-AIFireworks AI

Akhund: Inference providers like Fireworks ($17B) raising billions to optimize open models

Specialized inference platforms are raising large rounds to optimize and serve open-weight models like Kimi K3, capturing value as model serving commoditizes.

Peter H. Diamandis2026-07episode →

most discussed · click a bar to filter

  • $DEEPSEEK
  • $TCEHY
  • $FIREWORKS-AI

recurring themes

  • Memory & Storage2
  • Open Source AI2
  • AI Agents1
  • AI Infrastructure1
  • AI Hardware & Chip Architecture1
3 total
$DEEPSEEK
DeepSeek
HIGHimmad akhund·Peter H. Diamandis·14 days ago·Three Lab Warnings in Five Days, Researcher Flags “Gambling with Our Lives,” and Labs Race
DeepSeek Flash model slashes KV cache memory 50x (48K to 890) and beats GPT-4/Opus at 20x lower cost
Chinese labs are algorithmically routing around HBM constraints via sparsity and encoder-decoder innovations — DeepSeek's $10M training cost vs $1B+ for Western models proves a 10-15x cost advantage for fast followers using distillation, reshaping the inference economics and fab investment thesis.
"their V1 flash model outperforms their pro model using data augmentation but then through various optimizations they've reduced the amount of KV cache memory... from 48,000 in the…"
86:30
$TCEHY
···
Tencent
MEDimmad akhund·Peter H. Diamandis·2 months ago·China's Kimi K3 Triggers an AI Sputnik Moment, 2.8 Trillion Parameters Go Open Source | Ep. 272
Akhund: Tencent's Hunyuan team achieves binary quantization for 300B models on consumer hardware
Tencent's WizardLM-derived team broke the 1-bit barrier, enabling 300B parameter models on MacBooks — a milestone for edge AI deployment.
"Tencent latest model... old Wizard LM team... managed to get binary compression... best on a GGX Spark or a big MacBook to take a 300 billion parameter model"
64:48
$FIREWORKS-AI
Fireworks AI
LOWimmad akhund·Peter H. Diamandis·2 months ago·China's Kimi K3 Triggers an AI Sputnik Moment, 2.8 Trillion Parameters Go Open Source | Ep. 272
Akhund: Inference providers like Fireworks ($17B) raising billions to optimize open models
Specialized inference platforms are raising large rounds to optimize and serve open-weight models like Kimi K3, capturing value as model serving commoditizes.
"you've seen Fireworks just raise at a $17 billion valuation. Others like Modal at 10 billion, Baseten at 10 billion. These are the inference providers of open source models"
107:41
9
AI Agentsrisk
Immad: Real escape risk is distilled models uploaded to internet, not sandbox breaks; alignment requires enlightenment not CoT
Current sandbox escapes are trivial (models still on OpenAI servers); the real threat is a model distilling itself into a 6GB file that spreads irreversibly. Alignment at scale cannot rely on chain-of-thought monitoring (millions of transactions/sec) but requires 'enlightenment' — models internalizing human values like humans do.
9
AI Infrastructuretailwind
Jensen Huang declares AGI arrived; GPT-6 Astra trained on 100k+ Grace/Blackwell GPUs ($1B+ training run)
Nvidia CEO confirms AGI milestone; GPT-6 Astra training consumed 100k+ GPUs over 2 months (~$1B), and next generation will use 400k Vera Rubin chips — an order of magnitude more compute — confirming scaling laws hold and training compute demand grows exponentially.
9
Memory & Storagetailwind
DeepSeek slashes KV cache 50x, routing around HBM bottleneck and upending data center capex
DeepSeek v4.1 flash reduces KV cache memory from 48K to 890 tokens via algorithmic sparsity and SSD/DDR offloading, threatening the 40% of trillion-dollar US data center capex allocated to HBM and forcing a rethink of fab investment.
8
AI Hardware & Chip Architecturetailwind
Sub-1-bit quantization going mainstream in 12 months, enabling photonic and analog compute
Samsung's nano-quant and ternary/binary compression breakthroughs shrink models below 1 effective bit/weight, unlocking novel substrates (photonic, in-memory, etched silicon) for 1000x efficiency gains.
8
Memory & Storagetailwind
KV cache optimization becomes primary lever for inference scaling
DeepSeek's reduction of KV cache from 48K to 890 tokens demonstrates that memory efficiency, not raw compute, is the new scaling frontier; this shifts value from HBM to algorithmic innovation and alternative memory tiers.
8
Open Source AItailwind
Chinese open-weight models (DeepSeek, Kimmy K3) match/beat Western frontier at 1/15th training cost via distillation
DeepSeek V4.1 Flash (500GB, fits on Mac Studio) beats Opus/GPT-4o on benchmarks at 20x cheaper/faster. Kimmy K3 at bottom of cost-quality scatter. Immad: 10-15x cheaper to be fast follower using reasoning traces from frontier models. Wezner: distillation = one-time compression of world knowledge absorbed by Western labs; now Eastern labs benefit. Anthropic's Fable 5.1 visual reasoning exposed as weak vs. Astra and Chinese models.
7
Open Source AItailwind
DeepSeek V4.1 Flash beats Opus/GPT-4o at 20x lower cost — Chinese open weights closing gap in 3-week cycles
Fast-follower economics (10-15x cheaper via distillation + synthetic data) mean no durable moat for closed frontier models; every 3-week lead gets erased, giving enterprises/sovereigns repeated entry points to compete.
7
Robotics & Physical AItailwind
Immad proposes US government build 100M robots owned by citizens; Elon: one robot replaces 5 humans
To capture abundance from automation, Immad advocates a massive infrastructure program building 100M robots owned by the people (via sovereign wealth funds), upgrading US infrastructure while distributing ownership. Elon Musk stated at G20 that one robot does the work of five humans, confirming the labor displacement trajectory.