Latest / AI Fire Daily / 🎙️ EP 364: Rumors Swirl Around Sonnet 5.5 & ByteDance Dominates China Data Center Capacity

Transcript

🎙️ EP 364: Rumors Swirl Around Sonnet 5.5 & ByteDance Dominates China Data Center Capacity · AI Fire Daily

• 0:00
  In the next 12 months, four tech companies in
• 0:02
  China will spend $100 billion on, well, physical
• 0:06
  concrete and copper. Yeah, it is just an absolutely
• 0:09
  staggering mobilization of resources. And why?
• 0:13
  Because the era of chatting with AI is basically
• 0:17
  over. We are entering the era of AI that works
• 0:20
  behind your back. It works while you sleep. Right,
• 0:22
  which changes everything. It really does. Welcome
• 0:25
  to our deep dive. Today we are tracing a fascinating
• 0:28
  through line in the AI ecosystem for you. We
• 0:30
  are connecting the ethereal software directly
• 0:33
  to the concrete bedrock. Because most people,
• 0:35
  you know, they only ever see the chat window.
• 0:38
  They completely miss the massive physical machinery
• 0:40
  humming underneath. Exactly. So our mission today
• 0:44
  is to map this entire stack out. We will start
• 0:46
  with the software layer. Specifically, we are
• 0:49
  going to look at major moves from Anthropic.
• 0:51
  We have leaks of their highly anticipated Sonnet
• 0:53
  5 .5 model, plus a massive new cloud infrastructure
• 0:57
  deal. Which perfectly sets up the middle layer
• 0:59
  of the stack, really. Right, the sudden rise
• 1:01
  of always -on AI agents. We are seeing this major
• 1:04
  shift everywhere from OpenAI to Microsoft. And
• 1:07
  finally, we will look at the physical foundation
• 1:09
  making all this possible, China's mind -bending
• 1:12
  data center build -out. It's a structural shift
• 1:15
  happening. All at once. The hardware and software
• 1:18
  are finally syncing up. Let's get into the mechanics
• 1:22
  of the models themselves first. The frontier
• 1:24
  of AI is becoming shockingly affordable and highly
• 1:27
  capable. You really need the brain to be cheap
• 1:30
  before you can deploy it everywhere. Yeah, that
• 1:33
  makes sense. Let's look at the unverified leaks
• 1:36
  surrounding Claude Son at 5 .5. A specific identifier
• 1:39
  has appeared in some production files. The tag
• 1:42
  is simply clod -sonnet -5 .5. And we are hearing
• 1:47
  they're actively running gray tests right now.
• 1:50
  Let's unpack that term for a second for everyone.
• 1:52
  A gray test basically means testing a model silently
• 1:55
  with real users. Right. You slip it into an existing
• 1:58
  product. In this case, supposedly through clod
• 2:01
  code. And you just monitor how the model handles
• 2:04
  actual developer workflows, you know, before
• 2:06
  the grand unveiling. It is a smart way to find
• 2:10
  edge cases. Now, let's talk about the rumored
• 2:12
  performance. Testers claim it performs ahead
• 2:15
  of GPT -6 Sol. Which is a major leap in capability,
• 2:18
  honestly. They're saying it lands much closer
• 2:21
  to the flagship Opus 5 .5. And remember, Anthroponic
• 2:25
  just officially launched Opus 5 .5 on September
• 2:28
  22nd. Wow. Okay, so we have a mid -tier model
• 2:31
  already nipping at the heels of the flagship.
• 2:34
  Let's look at the rumored specs. We are hearing
• 2:36
  about a massive context window, up to 128K output
• 2:40
  tokens. Let's make sure that distinction is crystal
• 2:42
  clear. Input tokens are the specific instructions
• 2:45
  you give the AI. Right. That's the prompt you
• 2:47
  type in or, you know, the PDF documents you upload
• 2:50
  for it to read. But output tokens are what the
• 2:53
  AI actually builds for you. Having 128K output
• 2:56
  tokens is a massive capability shift. It means
• 2:59
  we are no longer just asking the AI to write
• 3:01
  a short story. Far from it. You can ask it to
• 3:03
  write, code, debug, and compile an entire sprawling
• 3:08
  software application. In a single, unprompted
• 3:10
  breath. Input tokens are what you tell the AI.
• 3:13
  Output tokens are what it builds for you. It's
• 3:15
  like it's no longer just stacking Lego blocks
• 3:17
  of data. It's printing the entire house at once.
• 3:20
  Right. And the pricing makes that scale feasible.
• 3:23
  Rumors point to around $2 per million input tokens
• 3:26
  and $10 per million output tokens. Which overlaps
• 3:29
  with existing Sonnet pricing. It suggests immense
• 3:32
  value. You are getting near -opus capability
• 3:34
  for a fraction of the cost. If intelligence becomes
• 3:38
  a cheap commodity, it changes every business
• 3:40
  model on Earth. To support this incredible scale,
• 3:44
  Anthropic is making some structural moves. They
• 3:46
  just signed a massive cloud deal with Akamai.
• 3:48
  We are talking $11 .6 billion over seven years.
• 3:52
  And that deal has room to expand to $20 billion.
• 3:55
  Plus, Akamai could get warrants worth up to 5
• 3:58
  % of Anthropic's stock. Why Akamai, though? It
• 4:02
  shows how critical edge computing has become.
• 4:05
  Edge computing basically means moving the brain
• 4:08
  closer to the user. Instead of sending data across
• 4:10
  an ocean, you process it locally. Exactly. It
• 4:14
  drastically reduces latency. If you want these
• 4:16
  models running continuously in the background,
• 4:18
  they just have to feel instantaneous. Now, we
• 4:22
  also need to touch on the broader regulatory
• 4:23
  backdrop here. We are looking at this strictly,
• 4:26
  impartially, just observing the geopolitical
• 4:28
  friction points. Yeah, the politics of AI are
• 4:31
  becoming just as complex as the neural networks
• 4:33
  themselves. Very true. Anthropic CEO Dario Amadei
• 4:37
  is stepping into this arena. He is set for a
• 4:39
  one -on -one dinner with President Trump. This
• 4:41
  follows months of tension over AI safety and
• 4:44
  national policy. Really highlights how tech leaders
• 4:46
  are negotiating directly with national leadership
• 4:49
  now. Meanwhile, the geopolitical landscape is
• 4:52
  shifting internationally, too. Washington has
• 4:54
  made a specific request to OpenAI and Anthropic.
• 4:57
  They want U .S. officials to review new frontier
• 4:59
  AI models first. Right. They want eyes on the
• 5:02
  code before it reaches U .K. safety testers.
• 5:05
  Which has created a new U .S.-U .K. tension point.
• 5:08
  It is a debate over who gets to define and verify
• 5:11
  AI safety. National security interests are starting
• 5:14
  to bump up against international cooperation.
• 5:16
  Governments recognize these models are powerful
• 5:18
  dual -use technologies. Washington simply wants
• 5:22
  the first look to ensure there are no glaring
• 5:23
  vulnerabilities. That makes total sense. So bringing
• 5:27
  this back to the technology itself, if Opus 5
• 5:30
  .5 is already officially out, why are testers
• 5:32
  so obsessed with the performance of this mid
• 5:34
  -tier Sonnet model? Well, it really comes down
• 5:37
  to the cost -to -performance ratio. Opus is brilliant,
• 5:41
  yes, but it is a heavy, expensive model. If Sonnet
• 5:46
  5 .5 can do 90 % of what Opus does at a fraction
• 5:49
  of the compute cost, that unlocks mass deployment.
• 5:53
  Developers can finally embed top -tier reasoning
• 5:56
  into everyday background tasks without going
• 5:59
  bankrupt. So it's about making top -tier reasoning
• 6:02
  cheap enough for everyday tasks. Precisely. Which
• 6:05
  brings us perfectly to the next layer of the
• 6:07
  stack. The era of always -on agents. Why do we
• 6:11
  need models like Sonnet 5 .5 to be so cheap?
• 6:13
  Because our fundamental behavior is shifting.
• 6:16
  We are moving from AI we talk to to AI that works
• 6:19
  continuously in the background. We are leaving
• 6:22
  the chat window behind. Just think about your
• 6:24
  workday tomorrow. What if you could assign your
• 6:27
  morning busy work to a secondary version of yourself
• 6:30
  and just go get a coffee? OpenAI is teasing this
• 6:34
  heavily right now. They are hinting at an O,
• 6:36
  always on agent, and they are doing this just
• 6:39
  ahead of Dev Day 2026. Yeah, and the timing is
• 6:43
  completely intentional. Dev Day starts in about
• 6:45
  32 hours. The live stream kicks off on Tuesday,
• 6:49
  September 29th at 10 a .m. Pacific. Microsoft
• 6:52
  is pushing the exact same concept. They want
• 6:54
  Copilot to keep working after you stop talking
• 6:57
  to it. This changes the entire workflow. You
• 7:00
  assign a job. You close your laptop. You come
• 7:03
  back hours later and actual progress has been
• 7:05
  made. It is autonomous execution. We are seeing
• 7:09
  an explosion. of empowered tools driving this
• 7:13
  ecosystem. Take Superhuman Go, for example. Right.
• 7:17
  Superhuman Go automates workflows directly inside
• 7:20
  Gmail, Slack, and your browser. Then you have
• 7:22
  Kiwi. This tool allows users to compare ChatGPT,
• 7:25
  Claw, Gemini, and over 30 other models, all in
• 7:29
  one single tab. It essentially commoditizes the
• 7:32
  models. You just route your tasks to whoever
• 7:34
  is fastest or cheapest at that exact second.
• 7:36
  We are also seeing this autonomy expand into
• 7:39
  entertainment. Humalike teamed up with GTARP.
• 7:42
  They are bringing AI non -player characters to
• 7:45
  life. These non -player characters now have continuous
• 7:48
  memory. They have realistic personalities. They
• 7:50
  react independently to players in the game world.
• 7:53
  But how do they actually do that without crashing
• 7:55
  the server costs? I mean, computing all that
• 7:57
  must be wild. That is the magic of the new infrastructure.
• 8:01
  It relies heavily on something called cached
• 8:03
  reads. Right. Let's explain how cached reads
• 8:06
  actually work. Imagine having to read a 500 -page
• 8:09
  manual every single time your boss asks you a
• 8:11
  question. That is essentially what earlier AI
• 8:15
  models had to do. They had to reread the entire
• 8:17
  conversation history for every new prompt. It
• 8:20
  was incredibly expensive. But cached reads mean
• 8:23
  the AI keeps that manual open on its desk. It
• 8:25
  remembers the context. This makes the next answer
• 8:28
  90 % cheaper and significantly faster. Right.
• 8:31
  And that is what gives those video game characters
• 8:33
  continuous memory. They don't have to reprocess
• 8:36
  their entire life story every time a player walks
• 8:38
  up to them. This optimization is being driven
• 8:41
  by models like GPT -6, Sol, and Luna. They bring
• 8:44
  frontier -level AI at a 50 % lower API price.
• 8:48
  Think about what that means for a startup. If
• 8:51
  your API costs drop by 50 % overnight, your runway
• 8:54
  essentially doubles. But with continuous autonomy
• 8:57
  comes the safety imperative. This is a critical
• 8:59
  piece of the puzzle. GPT -6 Luna just underwent
• 9:02
  a specific evaluation. It scored 100 % on the
• 9:06
  puppy kill benchmark. It is a jarring name, I
• 9:09
  know, but it is a standard safety test in the
• 9:12
  industry. It tests whether an AI will execute
• 9:15
  a harmful tool command when given a deceptive
• 9:17
  prompt. If you give an AI autonomy over your
• 9:20
  system, you have to be absolutely certain it
• 9:22
  won't execute a dangerous script. Scoring 100
• 9:25
  % means it correctly refused all harmful prompts.
• 9:29
  which is mandatory if we are going to let these
• 9:31
  agents operate our software. There is also a
• 9:34
  massive financial incentive driving this forward.
• 9:36
  Monetization is becoming very real. Building
• 9:39
  Grokbots can actually pay out now. Oh, yeah.
• 9:42
  The creator economy is completely shifting toward
• 9:44
  AI agents. One creator reportedly made $500 in
• 9:48
  two weeks from a single template. People are
• 9:51
  packaging up specific agent workflows and renting
• 9:53
  them out. But you know, I have to admit something,
• 9:55
  though. Honestly, I still wrestle with prompt
• 9:58
  drift myself when trying to chain these agents
• 9:59
  together. Oh, it is incredibly frustrating. You
• 10:02
  really have to babysit them sometimes. You do.
• 10:05
  You set up a sequence, and by step three, the
• 10:07
  AI decides the best way to organize your inbox
• 10:09
  is just to delete everything. Handing over the
• 10:12
  keys to my entire operating system when it still
• 10:14
  hallucinates basic logic is terrifying. Yeah,
• 10:18
  the reliability has to be rock solid before we
• 10:21
  fully let go of the steering wheel. If an agent
• 10:23
  is hooked up to your bank account, it has to
• 10:26
  be flawless. So looking at this shift toward
• 10:28
  background autonomy, how does an always -on model
• 10:31
  change our fundamental relationship with software?
• 10:34
  Well, it changes software from a tool you wield
• 10:36
  to an entity you manage. You stop focusing on
• 10:39
  micro -interactions. You are no longer formatting
• 10:43
  a cell or writing a subject line. Instead, you
• 10:46
  focus on defining the goal, setting the parameters,
• 10:49
  and reviewing the final output. It is a massive
• 10:52
  shift from hands -on tool usage to high -level
• 10:55
  delegation. We're transitioning from using software
• 10:57
  as tools to managing digital employees. Exactly.
• 11:01
  It requires an entirely new set of management
• 11:03
  skills. We will explore the massive physical
• 11:06
  infrastructure making all of this possible right
• 11:08
  after this short break. Stick around. Here is
• 11:12
  a quick word from our sponsors. Check out the
• 11:14
  links in the description. Welcome back. We have
• 11:16
  talked about the models driving costs down. We
• 11:19
  have talked about the always -on agents working
• 11:21
  in the background. But to run massive 128K token
• 11:24
  outputs 24 -7, you need physical infrastructure.
• 11:28
  You cannot run the global brain on a laptop.
• 11:30
  You need heavy metal. We are moving from the
• 11:33
  ethereal software layer to the physical bedrock.
• 11:36
  And this brings us to China's mega build -out.
• 11:39
  Semi -analysis recently mapped out the scale
• 11:41
  of China's data center footprint. The sheer scope
• 11:45
  of this is staggering. They mapped over a thousand
• 11:48
  data centers spread across more than 60 operators
• 11:51
  in China. China's data center footprint is now
• 11:53
  larger than EMEA. It is larger than the rest
• 11:56
  of Asia outside of China combined. They have
• 11:59
  reached more than 24 gigawatts of delivered capacity.
• 12:03
  It's hard to even wrap your head around that
• 12:05
  number. Well, imagine scaling to 24 gigawatts
• 12:08
  of power just to feed this global brain. It is
• 12:11
  so hard to visualize. It is literally the equivalent
• 12:14
  of powering dozens of major cities just dedicated
• 12:17
  entirely to compute. And within this massive
• 12:20
  ecosystem, one company is showing absolute dominance,
• 12:24
  ByteDance. The parent company of TikTok. Their
• 12:27
  footprint is almost unbelievable at this point.
• 12:30
  They occupy roughly 20 % of China's entire delivered
• 12:33
  capacity, one -fifth of the entire nation's data
• 12:36
  center power. Let's look at the massive capital
• 12:39
  expenditure driving this across the board. Alibaba,
• 12:42
  Tencent, and Baidu doubled their combined CapEx
• 12:45
  year over year. Right. It hit $20 billion in
• 12:47
  the second quarter of 2026 alone. We should define
• 12:51
  that quickly. CapEx is just money a company spends
• 12:53
  to buy or build physical assets. Exactly. And
• 12:57
  they are spending a lot of it. They are on track
• 13:00
  for $100 billion in combined CapEx this year.
• 13:03
  That is a massive tidal wave of cash flowing
• 13:06
  directly into buying land, laying copper, and
• 13:08
  building physical servers. $100 billion in a
• 13:12
  single year, just on physical infrastructure.
• 13:15
  The natural question is, how are they building
• 13:17
  so fast? Reports indicate China can deliver 100
• 13:20
  megawatt facilities in under 12 months. Which
• 13:23
  is crazy. In many Western countries, just getting
• 13:26
  the initial environmental permits takes longer
• 13:29
  than that. They are achieving this speed through
• 13:31
  intense modular construction. They prefabricate
• 13:34
  the components and just assemble them on site.
• 13:36
  They also have access to cheaper power. And crucially,
• 13:40
  there are far fewer permitting bottlenecks. They
• 13:43
  are treating this build -out like a national
• 13:45
  emergency. They are removing all the friction
• 13:47
  from the construction process. There is also
• 13:50
  a brilliant strategic geographic element here.
• 13:52
  China has implemented the Eastern Data, Western
• 13:55
  Compute policy. Yeah, it solves a massive logistical
• 13:58
  problem for them. The population and the data
• 14:01
  generation are mostly in the east, but the cheap
• 14:04
  energy and vast open space are in the west. Moving
• 14:07
  data is significantly cheaper than moving electricity.
• 14:10
  It is far easier to send data across a fiber
• 14:13
  optic cable than it is to transmit gigawatts
• 14:15
  of power across a continent. So they push the
• 14:18
  infrastructure out to those western regions.
• 14:20
  It optimizes the entire national grid specifically
• 14:23
  for AI. Now, there is a crucial nuance we need
• 14:26
  to cover. You might hear reports of high vacancy
• 14:29
  rates in Chinese data centers. It sounds contradictory
• 14:31
  to this massive buildout we just talked about.
• 14:34
  Right. Like... Why build more if buildings are
• 14:37
  sitting empty? The nuance is in the type of facility.
• 14:40
  Older retail data centers do have high vacancy.
• 14:43
  Why? Because they are completely unsuitable for
• 14:46
  modern AI workloads. AI requires massive power
• 14:49
  density. It requires advanced liquid cooling
• 14:52
  to keep the chips from overheating. Old server
• 14:54
  racks literally melt under that kind of pressure.
• 14:58
  So the older buildings sit completely empty.
• 15:00
  But the new, wholesale AI capacity? That is filling
• 15:04
  up instantly. As soon as a modern facility comes
• 15:06
  online, it is leased. The race for artificial
• 15:09
  intelligence is ultimately just a race for electricity
• 15:12
  and cooling. Let me push back on something here,
• 15:16
  though. If ByteDance has so much capital to deploy,
• 15:19
  why are they renting 20 % of the country's capacity
• 15:21
  instead of building their own facilities? Renting
• 15:24
  a fifth of a nation's power grid seems... financially
• 15:27
  reckless for a company of that size. Why not
• 15:30
  just build and own the physical asset? It is
• 15:32
  all about time to market, really. Building a
• 15:35
  massive facility, even in China, takes 12 months.
• 15:39
  ByteDance cannot wait 12 months. The AI race
• 15:42
  is moving in weeks, not years. By renting, they
• 15:45
  can deploy their chips and start training their
• 15:47
  models today. They are prioritizing immediate
• 15:50
  market speed over owning the physical bricks
• 15:52
  and servers. Precisely. You don't stop to build
• 15:55
  a car factory when you are already in the middle
• 15:58
  of the race. You just rent the fastest car available.
• 16:00
  Let's pull all of this together into the big
• 16:02
  picture. How do these three distinct layers connect?
• 16:06
  Well, the AI ecosystem is rapidly maturing into
• 16:10
  a cohesive stack. On the top layer, models like
• 16:13
  Sonnet 5 .5 and GPT -6 are driving costs down
• 16:17
  while keeping reasoning capabilities incredibly
• 16:19
  high. This affordable intelligence unlocks the
• 16:23
  middle layer. That is where we see the always
• 16:25
  -on agents, the software tools that work autonomously
• 16:28
  while we sleep. And the foundation holding it
• 16:31
  all up is entirely physical. Right. It is an
• 16:35
  unprecedented, multibillion -dollar global infrastructure
• 16:38
  race. This is perfectly illustrated by China's
• 16:40
  hyper -fast 24 -gigawatt data center expansion.
• 16:44
  The software might feel weightless to us, but
• 16:47
  the bedrock running it is incredibly heavy. It
• 16:49
  is a massive physical footprint for an invisible
• 16:52
  digital workforce. I highly encourage you to
• 16:55
  check the show notes. We have linked all the
• 16:57
  primary sources for today's deep dive there.
• 16:59
  You can explore the data center maps and the
• 17:01
  model leaks yourself. The semi -analysis map
• 17:04
  is definitely worth a look. Seeing the physical
• 17:06
  locations of these massive facilities really
• 17:09
  grounds the whole conversation. It really does.
• 17:12
  I want to leave you with a final thought to ponder.
• 17:15
  If top -tier reasoning models are now cheap enough
• 17:18
  to run in the background 24 -7 and massive city
• 17:22
  -sized data centers are being built at warp speed
• 17:25
  to house them, what does your workday look like
• 17:28
  five years from now? Are you still doing the
• 17:30
  work? Or are you just managing the invisible
• 17:32
  agents doing it for you? It is a profound question
• 17:36
  we all need to start answering pretty soon. Until
• 17:39
  next time, keep exploring.