Deep Moat

Palantir doubled its revenue selling sovereign AI to corporations, and every company that ever shipped its data to a frontier lab is now wondering whether it funded its own competitor.

This week on Deep Moat:

◼ The Anthropic CTO sat on Figma's board and quit the day before a Figma competitor launched

◼ Send your know-how to a lab that still hasn't solved revenue, and they have every incentive to take the good parts

◼ Meta's Muse Glimmer goes agentic, Nvidia ships Nemotron 3.5 Lightning, Liquid AI ships a local vision model

◼ Why 32GB stopped being enough and hundreds of gigabytes of RAM becomes the professional norm

◼ Apple nearly doubled RAM upgrade prices, and memory stays expensive for at least two years

◼ There isn't much money in models, which is exactly why Google and Meta can afford to open source them

◼ Fine tuning is back: same cost, but intelligence that overtakes the top tier

◼ Context engineering came full circle, models are now boxed into their context and stopped thinking outside it

◼ A 50-hour ad workflow that pays for itself, and the copy step that still breaks every time

◼ Deterministic beats probabilistic: give the agent 50 templates, not a blank canvas

◼ Stop image generating landing pages, compose them with code instead

◼ Treat your agent like an employee on day one, and run it like McDonald's runs a kitchen

Chapters:

0:00 Cold open

0:20 Palantir and the case for sovereign AI

0:48 When your AI partner becomes your competitor

2:04 Cost, latency, and the Chinese alternative

3:15 Meta's Muse Glimmer goes agentic

4:28 You're going to need a lot more RAM

6:24 Nvidia's Nemotron 3.5 Lightning

8:17 Should Google open source Gemini?

8:54 There isn't much money in models

10:06 Liquid AI's local vision model

11:06 Fine tuning is making a comeback

14:35 The last 10% still needs a human

15:20 The 50-hour ad workflow

17:55 Why models are getting worse at copy

20:24 Deterministic beats probabilistic

23:04 Compose with code, don't image generate

24:41 Take AI out of what AI is bad at

26:41 Treat your agent like a new employee

28:24 AI is giving us more work, not less

29:21 Freestyle first, then build the workflow

Every benchmark keeps going up, and the copy keeps getting worse. So what exactly are the labs optimizing for, and is anyone still building for the actual work?

This week on Deep Moat:

◼ Why AI still can't make the logical jump that turns information into a new idea

◼ The context engineering trap: you feed it detail for inspiration, it treats it as a fence

◼ Fine tuning is quietly coming back, and a 40% intelligence lift at the same price is why

◼ Even at 99% accuracy, a task that runs 10,000 times a day fails 100 times a day

◼ Judge models that grade their own homework, and the intern problem nobody has solved

◼ OpenAI cuts 80% off the wrong model, while half of US startup tokens already come from Chinese open weights

◼ Cost per token is a vanity metric, cost per task is the one that bankrupts you

◼ A harness scored 95% on ARC-AGI-3 with Opus 5, up from 30%, so maybe the model was never the hard part

◼ MCP finally gets un-over-engineered under the Agentic AI Foundation

◼ OpenAI hits 1 billion users twice as fast as Facebook, and 0.3% of the world pays for any of it

Chapters:

0:00 Cold open

0:25 Why AI can't make logical jumps

2:42 Will fine tuning come back?

3:58 Are the models getting worse?

5:17 Encoding taste into prompts

9:17 The 99% accuracy problem

13:14 OpenAI cuts prices by 80%

15:18 Cost per token vs cost per task

18:31 Chinese open weight models

21:02 95% on ARC-AGI-3

23:29 Bench maxing and over-optimization

24:57 MCP gets a rebuild

27:25 OpenAI hits 1 billion users

30:31 DeepMind's CEO steps back

Opus 5 finally gives us Fable-level performance without taking all of our money. So why does every task still spin for 15 minutes and cost about the same as before?

This week on Deep Moat:

◼ Opus 5 eats fewer tokens than Fable 5, but more than the generation before it, so what does a task actually cost now

◼ Why front-end design is still the weak spot no model has fixed

◼ The planning-phase problem: ask for one module, get five touched

◼ 35 minutes for a minor enhancement, and why we're switching to Grok 4.5 by default

◼ Why xhigh effort can be worse than high, and overthinking is now a real failure mode

◼ Anthropic is shredding rare books to scan them, while telling everyone else not to distill

◼ The open source fight: Jensen and OpenAI on one side, Anthropic asking for restrictions

◼ Kimi K3 hands-on, and why "the model is not the moat" keeps getting more true

Chapters:

0:00 Intro

0:29 Opus 5 pricing and token burn

2:25 Front-end design is still bad

3:19 Planning workflow and scope creep

4:15 Waiting 15 minutes for one task

5:34 Overthinking and effort levels

6:14 Anthropic is shredding books

8:14 Should the US restrict open source

10:12 Why NVIDIA wants open weights

12:10 Kimi K3 hands-on

14:17 Grok 4.5 efficiency and the C rewrite

14:50 Will xAI open weight a model

Grok 4.5 puts out around four times fewer tokens than GPT 5.5 or Opus 4.8, which is most of why it lands at four times the price cut. So if SpaceX AI has the GPUs, the cursor team and the rocket scientists, what is actually still holding them back?

This week on Deep Moat:

◼ Why Grok 4.5's token efficiency, not its benchmarks, is the real story

◼ Musk fires the Grok team, brings in SpaceX engineers, and buys cursor for the coding data

◼ The trillion parameter roadmap: 1.5 now, then 2, 3, 6, and 10

◼ Why catching Anthropic and OpenAI probably waits for Grok 5

◼ The uncomfortable read that SpaceX overtaking the top two knocks Google out of the top three

◼ Google's dark times, again, and why Gemini 3.5 Flash shipped while Pro slipped eight months

◼ Why a Gmail agent doesn't need the smartest model, just good enough, fast and cheap

◼ SpaceX holds nearly half the AI chips rolled out globally, and Musk is talking about a Terafab in orbit

Chapters:

0:00 Intro

0:28 Grok 4.5 and the token drop

2:00 Cost and speed as positioning

3:57 Why Grok fell behind

4:44 Musk fires the team, buys cursor

6:23 Why cursor's data matters

8:12 The trillion parameter roadmap

9:22 Does this push Google out?

10:48 Google's dark times

14:56 Google bets on AI search

16:56 SpaceX's chip moat

17:50 Chips in space

Anthropic's new top model, Fable 5, is smarter than anything else out there, so they're locking parts of it down, pulling it out of subscriptions, and charging double for it. Is this the most powerful AI yet, or the moment AI starts splitting into haves and have-nots?

This week on Deep Moat:

◼ Why Fable 5 sits in a whole new tier above Opus, and caught OpenAI off guard

◼ The cybersecurity and biochem restrictions that have people calling it "too controlled"

◼ Fable 5's effort levels, and why low effort already beats Opus 4.8 at high

◼ Pulling Fable from subscriptions after June 22, and the high-risk bet on per-token pricing

◼ Uber and Microsoft capping how much AI each employee can spend

◼ Price per outcome vs price per token, and the Stripe rebuild that took 1 day instead of 2 months

◼ Where Fable still loses: GPT 5.5 wins the vending benchmark, cheaper and faster on coding

◼ The real worry: AI for the elite, two classes, and "getting our AI democracy from China"

Chapters:

0:00 Intro & Fable 5 release

1:56 Hands-on: one-shotting & design

3:08 A new tier above Opus

4:54 Switching to Codex & GPT 5.5

7:01 The cost problem

9:06 Price per outcome, not tokens

11:56 Picking the right model

13:06 Effort levels explained

14:02 Benchmarks: vending & coding

16:21 Is Anthropic gatekeeping AI?

Opus 4.8 is 61% cheaper than 4.7 and a bit smarter on most benchmarks, but it lost something on the way there. This week we break down what Anthropic traded for that price drop, why everyone in business suddenly talks about Claude instead of ChatGPT, and what Uber's $1,500/month per-employee AI cap signals about where enterprise spending is headed.

We also get into the new effort levels in Claude, whether Opus 4.x has hit a ceiling that only a new architecture can break, NVIDIA's Vera CPU built specifically for agentic workflows, and why "half your customers being AI agents" is closer than most businesses are planning for.

Chapters:

0:00 Intro & Opus 4.8 release

2:58 Why Claude lost its personality

4:19 Picking the right effort level

5:16 Has Opus 4.x hit a ceiling?

6:25 Claude is overtaking ChatGPT

8:49 Claude for work, ChatGPT for personal?

11:35 Uber's $1,500 AI spend cap

14:30 How to actually measure AI ROI

16:18 NVIDIA's Vera CPU for agents

17:21 Agent-to-agent payments are coming

Microsoft just killed Claude Code inside the company, and Nvidia, Meta, and Uber are all rethinking how much they let employees burn on AI coding. The "token maxing" era is ending fast.

In this episode of DeepMoat we break down what that means for how teams should actually use AI in 2026: Microsoft's pivot to GitHub Copilot, why Uber torched a full year of API budget in three months, and the real reason behind the "use 50% of your salary on AI" memos.

We also get into Opus 4.7 second impressions, where it finally clicks (design work, presentations, day-to-day vs GPT-5.5), and why Anthropic's stale training data still bites in coding work.

Then we go deep on AI agents: what they're actually useful for today, how to use the Buy Back Your Time framework to pick which tasks to automate, why evals matter before you trust an agent's output, and why low-stakes repetitive work is still the sweet spot until hallucination rates drop further.

If you're a founder, engineer, or operator trying to figure out where AI coding tools and AI agents actually drive ROI versus where they burn money, this one's for you.

Chapters

0:00 Intro

0:24 Microsoft kills Claude Code

1:27 Nvidia and Meta's token-maxing

4:08 Why Microsoft pushed Copilot

4:42 Opus 4.7 second impressions

7:43 Anthropic's stale training data

8:50 Why co-work still frustrates

10:00 Opus 4.7 vs GPT-5.5 for slides

13:30 AI agents: hype vs reality

15:35 Picking the right tasks to automate

18:56 Low-stakes is the sweet spot

#ClaudeCode #AIAgents #AICoding

Google just shipped enough at I/O to become the Apple of AI. We get into Gemini 3.5 Flash beating last year's Pro, why Spark is built so you never leave Google, and how Universal Cart quietly killed SEO.

In this episode:

• What Gemini 3.5 Flash actually changes (and why it costs more)
• Why Spark is the lock-in play, not the agent play
• Gemini Omni vs Seedance and why everyone misses the point
• Universal Cart and what's left of search traffic
• Why Karpathy picked Anthropic over everyone

Timestamps

00:00 Intro

00:26 Gemini 3.5 Flash beats the old Pro

04:22 Why every AI model is getting more expensive

06:41 Gemini Spark and the OpenClaw question

12:14 Gemini Omni vs Seedance

16:08 Universal Cart and the death of SEO

22:34 Karpathy joins Anthropic

AI is quietly eating ecommerce, and Shopify's running out of moat. We get into Anthropic's new Cloud for Business, why 96% of AI deployments still fail, and how Stripe just made it possible for an agent to spin up a whole stack from a single command.

In this episode:

• What Cloud for Business actually is (and whether it solves anything)
• Why forward deployed engineers became tech's hottest role
• The case for ditching Shopify and wipe-coding your storefront with Claude Code
• Stripe Projects and the rise of agent-native infrastructure
• Why MCP servers are about to be unavoidable

Timestamps

00:00 Claude for Business: what Anthropic just launched

04:25 Forward deployed engineers and the 96% failure rate

09:01 Is Shopify in real trouble?

12:46 What's left of Shopify's moat

15:41 Stripe Projects and agent-native infrastructure

19:25 Why MCP servers are about to be table stakes

Anthropic is pulling ahead, OpenAI is fighting back, and the AI agent wars are getting messy.

In this episode, we break down Anthropic’s enterprise-first strategy, why Claude Code became a serious growth engine, and how OpenAI’s Codex comeback could shift the race again.

We also dig into Anthropic’s surprising SpaceX compute deal, the growing pains behind OpenClaw, and why the future of AI work may not be one giant agent - but a smarter orchestration layer of specialized agents.

Takeaways

• Anthropic’s enterprise-first strategy is paying off
  Claude Code vs Codex is becoming the key developer battleground
• The AI market is moving toward “service as software”
  Anthropic’s SpaceX compute deal could ease Claude’s capacity issues
• OpenClaw’s instability may be pushing users toward Claude Code and Codex
• The future may be unified interfaces with specialized sub-agents

Chapters

00:00 Intro

00:22 Anthropic Pulls Ahead of OpenAI
07:05 Anthropic’s SpaceX Compute Deal
10:04 OpenClaw vs Claude Code
16:03 The Future of AI Agent UX