Latest / AI Fire Daily / 🎙️ EP 361: Closed AI Labs Trigger Price War & Alibaba Unveils Qwen4 Stack & Zhenwu Chip

Transcript

🎙️ EP 361: Closed AI Labs Trigger Price War & Alibaba Unveils Qwen4 Stack & Zhenwu Chip · AI Fire Daily

• 0:00
  Sometimes the world changes, you know, incredibly
• 0:02
  quietly. Yeah, you don't really hear a loud crash.
• 0:05
  Exactly. The fundamental ground beneath you simply
• 0:08
  shifts. And that is exactly what happened very
• 0:11
  recently. On the exact same day, two giants quietly
• 0:15
  moved. They slashed the fundamental cost of intelligence.
• 0:19
  It was a massive tectonic shift. Yeah. We were
• 0:21
  talking about open AI and anthropic here. Right.
• 0:24
  They both dropped major landscape altering updates
• 0:27
  and they did it simultaneously. So welcome to
• 0:30
  this deep dive into your curated sources. Today,
• 0:33
  we are exploring a vicious new AI price war.
• 0:37
  It is getting incredibly competitive out there.
• 0:39
  It really is. We are also exploring how hardware
• 0:41
  is pushing massive models into our pockets. Then
• 0:44
  a $350 million bet on data. And finally, Alibaba's
• 0:48
  staggering $53 billion plan. They want to build
• 0:51
  a fully autonomous AI ecosystem. It is an incredibly
• 0:54
  dense stack of information today. The sheer pace
• 0:57
  of these breakthroughs is staggering. It is.
• 0:59
  I sit down to work with familiar tools, but they
• 1:01
  behave completely differently than they did yesterday.
• 1:04
  I mean, I still wrestle with prompt drift myself.
• 1:06
  Yeah, that feeling of vertigo is completely natural.
• 1:09
  It can be overwhelming. But that is exactly why
• 1:11
  we are here today. We're going to make sense
• 1:14
  of this chaos. So let's look at the cloud APIs
• 1:16
  first. The baseline cost of cognitive labor just
• 1:20
  plummeted. OpenAI and Anthropic launched frontier
• 1:24
  models on the exact same day. It absolutely feels
• 1:27
  like a highly coordinated Cold War. OpenAI came
• 1:30
  out swinging with GPT -6 Luna. I saw those numbers.
• 1:34
  They are wild. Totally wild. It sits at $0 .10
• 1:37
  per million input tokens. And it costs $0 .50
• 1:40
  per million for output. Which is, what, half
• 1:42
  the cost of the previous version? Exactly. Half
• 1:45
  the cost of GPT -5 .6 Luna. Wow. It totally changes
• 1:48
  the math for software developers. They also launched
• 1:50
  the more powerful GPT -6 Sol. How does that one
• 1:54
  price out? It costs $2 per million input tokens
• 1:56
  and $10 per million output tokens. It offers
• 2:00
  really strong performance at a much lower price.
• 2:02
  So it bridges the gap between budget and flagship.
• 2:05
  Right. And then you have Anthropic Strategic
• 2:07
  Countermove. Cloud Opus 5 .5 dropped its pricing
• 2:10
  by 20%. Okay, so what is their exact rate now?
• 2:14
  It's $4 per million input tokens. and $20 per
• 2:18
  million for the output. That is a serious reduction.
• 2:22
  Yeah, and the performance is incredibly close
• 2:24
  to the higher -end Fable 5 .1, but it is 40 %
• 2:27
  cheaper to run than Opus 5. I've been thinking
• 2:29
  about what this actually means fundamentally.
• 2:32
  It's like electricity becoming remarkably cheap
• 2:35
  a century ago. Suddenly, household appliances
• 2:38
  became viable for everyday people. That is the
• 2:41
  perfect historical analogy for this shift. Competition
• 2:44
  is no longer about having the smartest single
• 2:46
  model. Right. The new metric is simply intelligence
• 2:49
  per dollar. Exactly. It's about coding performance
• 2:52
  and running millions of calls at scale. The open
• 2:54
  weight models and cheaper competitors are really
• 2:57
  forcing this pressure. Closed AI providers basically
• 2:59
  have no other choice left. None. Anthropic is
• 3:02
  also trying some fascinating ecosystem plays
• 3:04
  here. They released an updated developer playbook
• 3:07
  for Opus 5 .5 and Clawed Code. They are actively
• 3:10
  teaching people how to maximize efficiency. Yeah,
• 3:13
  and they are leaning heavily into cached tokens
• 3:16
  for long agent workflows. Let me just define
• 3:18
  that jargon quickly for everyone. Cached tokens
• 3:21
  mean saving previous texts to make responses
• 3:23
  cheaper and faster. Right, which encourages highly
• 3:26
  complex, multi -step agents. Anthropic even added
• 3:30
  a new usage reset button on web and desktop.
• 3:33
  Just to instantly refresh your banked limits.
• 3:35
  Exactly. It completely removes the friction of
• 3:38
  testing new models. But looking at these constantly
• 3:41
  shrinking margins, is this race to the bottom
• 3:44
  actually sustainable for closed providers? Well,
• 3:48
  they are trading immediate profits for total
• 3:50
  ecosystem capture. If they remain expensive,
• 3:53
  developers just migrate elsewhere. So cheaper
• 3:56
  models unlock massive scale, replacing high margins
• 3:58
  with high volume. Exactly. They want to become
• 4:01
  the invisible oxygen of software. So if the cloud
• 4:04
  is suddenly this incredibly cheap... Why pay
• 4:06
  for the cloud at all anymore? Your actual device
• 4:09
  can do the heavy lifting natively. The center
• 4:11
  of gravity is definitely moving to the edge.
• 4:13
  Qualcomm just launched two new smartphone chips
• 4:16
  focused on AI features. The technical specs on
• 4:19
  the premium model are fascinating. They are.
• 4:21
  It can run a 30 billion parameter Moe model completely
• 4:24
  locally. And just to clarify that term, Moe means
• 4:27
  routing tasks to specialized mini brains to save
• 4:31
  power. Exactly. You aren't firing up the whole
• 4:33
  engine every single time. Google is also pushing
• 4:35
  hard into this local integration. The new Google
• 4:38
  book is up for preorder now. Yeah, they are baking
• 4:41
  Gemini directly into the core workflow. It handles
• 4:43
  complex screen actions natively. It handles light
• 4:46
  transcription too, right? Yes. And deep Android
• 4:49
  connections without constantly uploading your
• 4:51
  data. But I think we need a serious reality check
• 4:53
  here. Oh, definitely. Promising magical AI features
• 4:56
  and actually delivering them are very different
• 4:58
  things. Look at the recent Apple news. Yeah,
• 5:00
  that was a brutal reminder for the industry.
• 5:02
  They are paying a $250 million settlement. It
• 5:06
  is specifically over delayed Siri AI feature
• 5:09
  claims for qualifying iPhones. Marketing departments
• 5:12
  consistently outpace the actual engineering realities.
• 5:15
  When you promise magic on a device, you simply
• 5:18
  have to deliver. Which brings up a very practical
• 5:21
  physical constraint. How will running local 30
• 5:25
  billion parameter models not instantly drain
• 5:27
  our laptop and phone batteries? Well, it requires
• 5:30
  a complete rethinking of the silicon architecture.
• 5:33
  These new chips have highly efficient, dedicated
• 5:37
  neural processing units. They crunch the AI math
• 5:41
  without waking up the main power -hungry processor.
• 5:44
  Right. Specialized silicon makes local processing
• 5:47
  far more efficient than older chips. It fundamentally
• 5:49
  changes the thermal and power dynamics of your
• 5:52
  phone. Whether you run models locally or in the
• 5:54
  cloud, though, they all desperately need two
• 5:56
  things. Massive amounts of training data and
• 5:59
  tools to make them usable. The hunger for specialized
• 6:01
  data is practically violent right now. Look at
• 6:05
  the sheer economics of it. Snorkel AI just raised
• 6:07
  $350 million. Yeah, their new valuation sits
• 6:10
  at $3 .5 billion. That is nearly three times
• 6:12
  higher than 17 months ago. But the underlying
• 6:15
  revenue metric is what really caught my attention.
• 6:18
  Their revenue run rate jumped 18 times. 18 times.
• 6:21
  It is unbelievable growth. It makes complete
• 6:24
  sense when you visualize it. It's like stacking
• 6:27
  Lego blocks of data. You need perfectly structured,
• 6:30
  interlocking pieces of specific information.
• 6:33
  That 18 times revenue jump just screams absolute
• 6:38
  market desperation. The public Internet is basically
• 6:40
  tapped out. Companies must construct their own
• 6:43
  highly specific internal data sets now. This
• 6:46
  friction is causing massive shifts everywhere.
• 6:49
  The dynamic between OpenClaw and Meta's Muse
• 6:52
  is a perfect example. Yeah, OpenClaw went completely
• 6:55
  viral recently. People started calling it OpenClaw
• 6:57
  for normies. It became the accessible tool for
• 6:59
  everyone. And then Meta released their highly
• 7:01
  polished enterprise product, Muse. Meta insisted
• 7:04
  they built Muse entirely from the ground up originally.
• 7:08
  But they eventually admitted OpenClaw heavily
• 7:10
  shaped their product. Even though they built
• 7:12
  it themselves. This is a fascinating tension.
• 7:14
  But we are seeing incredible specialized tools
• 7:16
  pop up everywhere now. Look at what Cluso MCP
• 7:19
  is doing for video editing. That tool feels like
• 7:22
  a glimpse into tomorrow. You literally just chat
• 7:25
  with Claude or ChatGPT to edit videos. But the
• 7:28
  crucial detail is that it keeps everything completely
• 7:31
  editable. You aren't just generating a flat,
• 7:34
  unchangeable video file. Right. It is a non -destructive
• 7:38
  workflow. Then you have enterprise tools like
• 7:40
  Anomalo Analyst. That targets massive corporate
• 7:43
  data environments. It constantly monitors Snowflake,
• 7:46
  Databricks, and BigQuery for hidden trends. It
• 7:49
  highlights anomalies before you even ask the
• 7:52
  system a question. We are moving from reactive
• 7:54
  prompting... to proactive intelligence. But seeing
• 7:57
  all this automation makes me wonder, does the
• 7:59
  explosion of automated data tools mean we are
• 8:01
  running out of human data? Yes, we have largely
• 8:04
  exhausted the easily scraped human knowledge.
• 8:07
  These tools are now synthesizing new, incredibly
• 8:09
  complex training environments from scratch. Got
• 8:12
  it. They are generating highly specialized data,
• 8:14
  not just scraping the web. We are firmly entering
• 8:16
  the era of pure synthetic generation. Midroll
• 8:19
  sponsor, Reed Placeholder. So we've been analyzing
• 8:22
  these specific isolated Western battles. We looked
• 8:25
  at API pricing wars and local edge chips. The
• 8:28
  Western ecosystem is incredibly fragmented. Right.
• 8:30
  But in China, Alibaba is taking a radically different
• 8:33
  approach. They are attempting to own that entire
• 8:36
  pipeline at once. They are building massive,
• 8:38
  integrated, big stack energy. They want to be
• 8:42
  the ultimate full stack infrastructure provider.
• 8:45
  They are building a complete AI ecosystem to
• 8:48
  push into the global race. Let's start by looking
• 8:51
  at their foundational software. Quen4 is currently
• 8:54
  in heavy development. But their internal roadmap
• 8:56
  already stretches much further. They're actively
• 8:59
  designing Quen4 .5 and Quen5. And the scale they
• 9:02
  are aiming for is genuinely breathtaking. They
• 9:05
  are targeting 5 to 10 trillion parameters. That
• 9:08
  represents an astronomical leap in raw computational
• 9:11
  density. You cannot run that on standard hardware.
• 9:14
  So they introduced the Zenmoo V900 AI chip. They
• 9:18
  claim it delivers three times the performance
• 9:20
  of their previous generation. But the networking
• 9:22
  scale is what caught me. Whoa, imagine scaling
• 9:24
  to a cluster of 500 ,000 chips. The sheer physical
• 9:27
  coordination required for that is terrifying.
• 9:29
  They are backing this vision with unimaginable
• 9:31
  resources. Alibaba is planning to invest $53
• 9:35
  billion into AI infrastructure. They want 20
• 9:38
  gigawatts of data center capacity by 2032. 20
• 9:41
  gigawatts. That is a massive amount of power.
• 9:43
  They need it because they are researching fully
• 9:46
  autonomous AI systems. They are chasing recursive
• 9:49
  self -improvement workflows, which basically
• 9:52
  means models testing and rewriting their own
• 9:55
  code to get smarter automatically. Two secs,
• 9:58
  silence. It is a very heavy concept to digest.
• 10:02
  But look at the full ecosystem vision they have.
• 10:05
  You have Quen Foundation models sitting at the
• 10:07
  base. Then T -Head chips providing the hardware.
• 10:10
  Plus Alibaba Cloud managing the compute. And
• 10:13
  agent platforms deploying the actual applications.
• 10:16
  It is vertical integration on an unprecedented
• 10:18
  scale. But we do have to remember, these are
• 10:20
  just projections right now. Trillion parameter
• 10:22
  targets and chip claims always look flawless
• 10:25
  on paper. Yeah, a lot of companies make massive
• 10:27
  claims that... Right. And the actual physical
• 10:30
  engineering to get there is incredibly messy.
• 10:32
  Even if they fall slightly short, the ambition
• 10:34
  is just staggering. It makes me question the
• 10:36
  architecture of innovation. Is this massive top
• 10:39
  -down integration a smarter approach than the
• 10:41
  fragmented U .S. ecosystem? Well, vertical integration
• 10:45
  offers raw efficiency. It removes hardware bottlenecks,
• 10:48
  but fragmented systems generally pivot much faster
• 10:51
  during unexpected disruptions. So vertical integration
• 10:54
  offers raw efficiency, but might sacrifice rapid
• 10:57
  grassroots innovation. We're watching two completely
• 11:00
  divergent philosophies collide globally. This
• 11:03
  has been a deeply fascinating deep dive today.
• 11:06
  We examined how the raw cost of intelligence
• 11:08
  is plummeting. Which fundamentally changes how
• 11:11
  developers design software. It allows AI to migrate
• 11:14
  directly onto our personal edge devices. Reducing
• 11:18
  our reliance on the cloud. Meanwhile, behind
• 11:20
  the scenes, infrastructure investments are scaling
• 11:22
  to the tens of billions. They are preparing for
• 11:25
  recursive, self -improving systems. The scale
• 11:28
  is just hard to wrap your head around. It leaves
• 11:30
  me with a provocative thought for you to chew
• 11:32
  on. If recursive self -improvement actually works
• 11:34
  at the 10 trillion parameter scale, what happens
• 11:37
  next? Will the next AI price war even be measured
• 11:41
  in dollars? Or will energy itself become the
• 11:44
  only currency that matters? It is a profound,
• 11:46
  slightly unsettling horizon approaching. Thank
• 11:49
  you for diving into these sources with us today.
• 11:51
  The landscape is shifting incredibly fast. We
• 11:54
  encourage you to test the new pricing limits
• 11:56
  on your own side projects. Build something weird
• 11:58
  and complex. Keep exploring these tools. Until
• 12:01
  next time.