Latest / AI Fire Daily / 🎙️ EP 387: Microsoft Hits $92B Record Quarter & Google’s Dangerous Gemini 4 Carbon!

Transcript

🎙️ EP 387: Microsoft Hits $92B Record Quarter & Google’s Dangerous Gemini 4 Carbon! · AI Fire Daily

• 0:00
  Tech giants are raising nearly $100 billion for
• 0:02
  infrastructure right now. At the exact same time,
• 0:06
  new AI models are being deemed too dangerous
• 0:08
  for public release. Welcome to this custom -tailored
• 0:12
  deep dive. Glad to be here. Yeah. We're unpacking
• 0:17
  a massive technological shift today. Just looking
• 0:20
  at this fresh... stack of sources reveals a rapidly
• 0:23
  changing landscape. Oh, absolutely. So here's
• 0:26
  our roadmap for you today. First, we'll explore
• 0:29
  the staggering financial scale of AI infrastructure,
• 0:32
  because Microsoft and Meta are quietly building
• 0:34
  physical empires. Next, we examine how that computing
• 0:38
  power is birthing new models, models that are
• 0:40
  now locked behind strict security clearances.
• 0:43
  We'll look at Gemini 4 Carbon specifically. Yeah,
• 0:45
  that one is fascinating. It really is. And we'll
• 0:48
  also discuss the political push to rebrand this
• 0:50
  tech as super intelligence. Finally, we'll track
• 0:54
  the everyday economics of this boom. We want
• 0:57
  to see exactly how it trickles down to your wallet.
• 0:59
  Right. And this stack of notes changes every
• 1:01
  single week. It's frankly, it's hard to keep
• 1:03
  up with the pace sometimes. It really is. So
• 1:06
  I want to start with the physical reality of
• 1:08
  this. All right. So Microsoft just posted its
• 1:10
  10th straight record quarter. 10th. Yeah. That
• 1:14
  is that's just wild. Yeah. Yeah. They hit $92
• 1:18
  .3 billion in revenue, and that's up 19 % year
• 1:23
  over year. The driver is cloud computing. That
• 1:27
  is absolutely the big driver here. Cloud is absolutely
• 1:29
  carrying the broader tech market right now. I
• 1:32
  mean, Microsoft Cloud brought in nearly $64 billion
• 1:35
  alone. Wow. Yeah, Azure specifically jumped 48
• 1:38
  % in revenue. It's not just a software product
• 1:41
  anymore. It's basically becoming the utility
• 1:44
  grid for the modern internet. That concept is
• 1:47
  fascinating to me. Like, we used to buy software
• 1:50
  in physical boxes. Right. Now a company is just
• 1:52
  rent processing power like electricity. And the
• 1:56
  backlog of contracted revenue really proves this
• 1:58
  shift. Their remaining performance obligation
• 2:00
  jumped 73%. Yeah, that number is huge. It sits
• 2:03
  at $677 billion. That backlog is entirely future
• 2:08
  revenue that's already locked in. Companies are
• 2:10
  terrified of running out of compute power. They're
• 2:12
  signing these massive contracts for years ahead.
• 2:15
  Because they have to. Exactly. They know they'll
• 2:18
  need these resources desperately. Without server
• 2:20
  space, enterprise AI initiatives just completely
• 2:22
  stall out. Well, imagine scaling to a billion
• 2:26
  queries. The physical reality of that is staggering.
• 2:29
  It really is. It's hard to wrap your head around.
• 2:31
  I'm trying to visualize this massive infrastructure
• 2:33
  build -out. It's like stacking Lego blocks of
• 2:36
  data centers across the globe. That's a perfect
• 2:38
  way to conceptualize it. They literally build...
• 2:41
  standardized modular server units now. They just
• 2:45
  snap them together wherever there's cheap power
• 2:48
  available. It's a massive physical earth moving
• 2:52
  operation. It requires sprawling facilities and
• 2:55
  totally dedicated power grids. The hardware side
• 2:58
  tells a completely different story though. Oh
• 3:00
  for sure. Traditional consumer computing is actually
• 3:02
  dying right now. Windows OEM and devices dropped
• 3:06
  22 percent. Yeah. That's a massive decline for
• 3:10
  traditional desktop computing. People are just
• 3:12
  holding on to their old laptops much longer.
• 3:14
  You really don't need a supercomputer on your
• 3:17
  desk anymore. Right, because of the cloud. Exactly.
• 3:20
  The heavy lifting happens miles away in those
• 3:22
  remote servers. Microsoft is betting everything
• 3:25
  on that remote capacity. And Meta's following
• 3:27
  the exact same playbook right now. They're raising
• 3:30
  $95 billion for AI infrastructure. $95 billion.
• 3:34
  That's essentially the GDP of a small nation.
• 3:38
  unbelievable amount of concentrated capital.
• 3:40
  And the chip makers are funding their biggest
• 3:42
  clients directly, too. Wait, really? Yeah. NVIDIA
• 3:45
  is putting $3 billion into that Meta raise, and
• 3:48
  AMD is investing another billion right on top
• 3:51
  of that. That feels like a massive strategic
• 3:53
  alignment between them. Like they're ensuring
• 3:55
  Meta stays locked into their specific hardware
• 3:58
  architecture. Oh, it's a complete arms race for
• 4:01
  computing dominance. You can't just buy standard
• 4:02
  servers for this stuff anymore. You have to build
• 4:05
  custom silicon clusters from scratch. You basically
• 4:08
  need dedicated power plants just for these facilities.
• 4:11
  With cloud booming and consumer hardware dropping
• 4:14
  22%, are our local computers basically becoming
• 4:18
  obsolete? Well, their core function is definitely
• 4:21
  shifting dramatically. You just need a fast,
• 4:24
  reliable internet connection. today right your
• 4:26
  device is basically just a screen and a keyboard
• 4:28
  the actual processing happens in an azure data
• 4:31
  center somewhere so local devices are just dumb
• 4:34
  windows into massive cloud brains precisely we're
• 4:37
  outsourcing our computing power to these massive
• 4:40
  tech giants and you know that staggering hardware
• 4:42
  spending isn't for standard enterprise software
• 4:45
  tasks no it's definitely not it's specifically
• 4:47
  required to train the next wave of intelligence
• 4:49
  that's the genuine causal connection we need
• 4:52
  to highlight here you don't build a 95 billion
• 4:55
  dollar data center for basic chatbots. Exactly.
• 4:58
  You build it to train something entirely new
• 5:00
  and autonomous. These next generation models
• 5:03
  are suddenly too powerful for the general public.
• 5:06
  The jump in sheer capability is happening incredibly
• 5:10
  fast. I mean, Google just introduced their newest
• 5:12
  model, Gemini 4 Carbon, and it follows Gemini
• 5:16
  4 Argon by a mere two weeks. Just two weeks.
• 5:19
  Yeah. The development cycles are compressing
• 5:21
  rapidly right before our eyes. Carbon is their
• 5:24
  newest frontier model in the Gemini lineup. And
• 5:27
  for those listening, a frontier model is, well,
• 5:30
  let's define it, the most advanced, cutting -edge
• 5:33
  AI systems currently available today. Right.
• 5:36
  It's the absolute bleeding edge of the technology.
• 5:38
  And Carbon is heavily focused on complex, autonomous,
• 5:43
  agentic work. It handles deep cybersecurity analysis
• 5:46
  and long -running, multi -step coding tasks perfectly.
• 5:51
  You can't actually use carbon right now. Google
• 5:54
  locked it tightly behind their internal Fairwind
• 5:56
  program. Yeah, they did. They're only rolling
• 5:58
  it out to trusted, vetted cyber defenders. They've
• 6:01
  essentially deemed it too dangerous for general
• 6:03
  public release. Wow. A system that understands
• 6:06
  software vulnerabilities this well is incredibly
• 6:09
  risky. Developer and consumer access is being
• 6:13
  heavily delayed by design. I want to look at
• 6:15
  the actual cybersecurity benchmarks here. Because
• 6:18
  it hit 88 .2 % on SecBench Pro. I'm trying to
• 6:23
  understand what that actually means in practice.
• 6:25
  Well, that benchmark tests finding real, deeply
• 6:28
  hidden software vulnerabilities. It scans massive,
• 6:31
  complex code bases to spot tiny, exploitable
• 6:34
  flaws. Oh, I see. Hitting near 90 % accuracy
• 6:37
  on that is just unheard of. It spots things human
• 6:40
  auditors easily miss. It also scores 71 % on
• 6:44
  CWE Bench. That specific benchmark tests producing
• 6:49
  working security patches, doesn't it? Exactly.
• 6:51
  So it doesn't just find the broken lock on the
• 6:54
  door. It designs a better lock and installs it
• 6:57
  completely autonomously. That's wild. It acts
• 7:00
  as a tireless, highly skilled senior cybersecurity
• 7:02
  engineer. That clearly explains why they locked
• 7:06
  it down so tightly. If it can fix vulnerabilities,
• 7:09
  it can obviously exploit them too. Oh, absolutely.
• 7:11
  It could write highly sophisticated malware in
• 7:14
  the wrong hands. It also reached 0 % attack success
• 7:18
  on Gray Swan. That's a highly rigorous benchmark
• 7:21
  for indirect prompt ejection. We should define
• 7:24
  that jargon so we're all entirely clear. Prompt
• 7:27
  ejection is tricking an AI into ignoring its
• 7:31
  original instructions to do harm. Indirect prompt
• 7:34
  injection is even scarier in practice. The malicious
• 7:37
  instructions are hidden away inside a random
• 7:39
  website. Right. The AI reads the site and gets
• 7:42
  hijacked from the inside. It's incredibly difficult
• 7:44
  to defend against. Which makes that 0 % figure
• 7:47
  pretty staggering. I still wrestle with prompt
• 7:50
  drift myself, so that precision is wild. Getting
• 7:52
  a 0 % failure rate is a monumental engineering
• 7:55
  achievement. It means Carmen is incredibly resilient
• 7:58
  against external manipulation attempts. Yeah.
• 8:00
  But running this massive capability isn't...
• 8:03
  cheap by any means. Not at all. It starts at
• 8:06
  $2 per million input tokens and it costs $10
• 8:09
  per million output tokens to generate. Okay.
• 8:12
  That actually matches the introductory pricing
• 8:14
  of their previous Argonne model. But Google says
• 8:17
  that price will eventually jump significantly
• 8:18
  higher. It will rise to $4 and $20 respectively.
• 8:22
  Yeah, that's a steep hike. Running complex autonomous
• 8:25
  workflows at that price gets expensive incredibly
• 8:28
  fast. And this immense power is also causing
• 8:31
  a broader political shift. We're seeing semantic
• 8:33
  changes reflect the technology's rapidly growing
• 8:36
  power. Right. And to be clear, we're just observing
• 8:39
  this linguistic shift impartially from the outside.
• 8:41
  Of course. Donald Trump is mandating a ban on
• 8:44
  the term AI. This applies to all relevant documents
• 8:48
  in his political sphere. He wants to push for
• 8:51
  the term superintelligence instead. He views
• 8:55
  these systems as far beyond mere artificial intelligence.
• 8:58
  It's a calculated push to rebrand the entire
• 9:00
  technological field. Yeah. The terminology we
• 9:03
  use ultimately shapes how we perceive the threat.
• 9:05
  Does changing the name to superintelligence actually
• 9:08
  change how we regulate it? The framing absolutely
• 9:11
  dictates the policy response in many ways. If
• 9:14
  you call it superintelligence, you regulate it
• 9:16
  heavily. You treat it more like a weapon than
• 9:19
  a software tool. Words matter immensely when
• 9:22
  lawmakers draft framework legislation. Right.
• 9:25
  AI feels too small for what these systems can
• 9:28
  actually do now. We're definitely outgrowing
• 9:30
  the traditional terminology of the past decade.
• 9:32
  Calling it artificial intelligence almost undersells
• 9:35
  the vast agentic capabilities. We need to take
• 9:38
  a quick pause here. When we come back, we'll
• 9:41
  look at everyday consumer economics. We'll see
• 9:43
  how this infrastructure boom trickles down to
• 9:45
  your wallet. We'll be right back. Sounds good.
• 9:48
  Sponsor? All right, let's get back into it. Welcome
• 9:51
  back to the Deep Dive. We just explored these
• 9:54
  massive, heavily guarded superintelligence models
• 9:56
  because they're so resource heavy. the everyday
• 9:59
  economics are drastically changing. The cost
• 10:03
  of running inference on frontier models is astronomically
• 10:06
  high. Companies are being forced to find highly
• 10:09
  creative monetization strategies. The free research
• 10:11
  phase is officially over for everyday developers.
• 10:15
  Anthropic is seeing this financial reality play
• 10:17
  out right now. Cloud spending is absolutely exploding
• 10:19
  for their enterprise and developer clients. It
• 10:22
  really is. So they launched brand new tools to
• 10:25
  track spending granularly. Teams can now see
• 10:28
  exactly what the AI produced each day. They calculate
• 10:30
  what those specific text generations cost in
• 10:33
  real time. Which is so necessary. Exactly. They
• 10:36
  can pinpoint exactly where a complex workflow
• 10:39
  wasted valuable tokens. Granular tracking is
• 10:42
  essential when API costs are this incredibly
• 10:45
  high. You can't just blindly query the model
• 10:47
  all day anymore. Nope. Meanwhile, OpenAI is trying
• 10:50
  a completely different tactic to retain users.
• 10:53
  They're offering eligible pro and business users
• 10:55
  massive API subsidies. They're handing out up
• 10:58
  to $500 a month. These are API credits just to
• 11:02
  keep developers building on ChatGPT. OpenAI wants
• 11:06
  to maintain their massive ecosystem lock -in
• 11:08
  at all costs. They know developers will jump
• 11:11
  to cheaper open -source models otherwise. Giving
• 11:14
  away compute credits builds long -term loyalty
• 11:15
  to their specific API. It's a smart play. Then
• 11:19
  we have Google's aggressive approach to consumer
• 11:20
  search monetization. Google just expanded ads
• 11:24
  directly inside its AI search worldwide. Core
• 11:27
  search revenue growth has slowed down in recent
• 11:30
  quarters. Their immediate answer is shoving sponsored
• 11:32
  links inside the AI answers. Generating an AI
• 11:36
  summary costs much more than a standard search.
• 11:39
  Google has to offset that massive inference cost
• 11:41
  somehow. It alters the clean visual aesthetic,
• 11:44
  but unit economics demand it. Placing ads directly
• 11:47
  into conversational output is very tricky, though.
• 11:51
  It has to feel natural without being overly deceptive
• 11:53
  to users. But Google's core business model demands
• 11:56
  search monetization above all else. With Google
• 11:59
  putting ads inside AI answers and API costs exploding,
• 12:03
  are we losing the pure utility phase of AI? Yes,
• 12:08
  the honeymoon phase of clean, ad -free AI is
• 12:11
  officially over. These companies spent nearly...
• 12:14
  $100 billion on physical infrastructure. Right.
• 12:17
  They have to show Wall Street a return on that
• 12:19
  massive investment soon. Basically, monetization
• 12:22
  always finds a way to infiltrate the clean answer
• 12:25
  stream. It's the iron law of consumer technology
• 12:28
  platforms at massive scale. But it's not all
• 12:32
  doom and gloom for everyday users. No, definitely
• 12:34
  not. We're seeing some incredibly practical everyday
• 12:37
  tools emerge right now. I want to dive into a
• 12:40
  few of those practical applications. A self -hosted
• 12:43
  job hunting tool is blowing up online lately.
• 12:45
  It helps you find open roles and tailor your
• 12:48
  resumes automatically. It even crafts custom
• 12:51
  cover letters based on the specific job description.
• 12:54
  That is helpful. You track every single application
• 12:57
  in one private centralized dashboard. It takes
• 13:00
  the tedious soul -crushing friction out of the
• 13:02
  modern job hunt. Being self -hosted is a massive
• 13:06
  benefit for user privacy, too. It means you control
• 13:08
  your own personal data completely. Exactly. Another
• 13:12
  brilliant example is a project by a creator named
• 13:14
  Emily Lai. She used AI to analyze 1 ,500 viral
• 13:18
  X replies. The replies were about products people
• 13:21
  genuinely can't live without. Reading unstructured
• 13:24
  data like that manually would take weeks of effort.
• 13:27
  Yeah, nobody has time for that. She used an AI
• 13:29
  model as a massive automated synthesis engine
• 13:31
  instead. She curated all of that chaos into one
• 13:34
  searchable website. Everything listed on her
• 13:37
  curated site is under $250. It's brilliant. It's
• 13:39
  a perfect example of transforming raw data into
• 13:42
  practical utility. It proves AI is incredible
• 13:45
  at finding signal in the noise. It took human
• 13:47
  chaos and structured it into a highly valuable
• 13:49
  resource. We're also seeing AI move aggressively
• 13:52
  into the physical world now. X -Rail Aura is
• 13:55
  a fascinating new augmented reality tool. It
• 13:57
  lets Gemini see your physical view through smart
• 14:00
  glasses. Yeah, that blew my mind. It merges digital
• 14:02
  intelligence with physical reality seamlessly.
• 14:04
  It can literally clean up your virtual windows
• 14:07
  in spatial computing. You can brainstorm with
• 14:09
  the AI based on whatever you're looking at. That's
• 14:11
  wild. The model uses the camera as its direct
• 14:14
  visual input stream. It's essentially giving
• 14:16
  the AI eyes in the real world. Finally, let's
• 14:19
  talk about some major voice AI advancements.
• 14:22
  Rhyme just released an important enterprise tool
• 14:24
  called Prism. Prism gives enterprise companies
• 14:26
  something voice AI badly needs right now. It
• 14:29
  provides a chance to check the words before they're
• 14:31
  spoken. Which is huge. It is. Hallucinations
• 14:34
  are incredibly dangerous when an AI speaks directly
• 14:37
  to customers. Oh, a total PR nightmare. Prism
• 14:40
  lets companies double -check the text output
• 14:42
  for complete accuracy. It acts as a mandatory
• 14:46
  safety checkpoint before audio is synthesized.
• 14:49
  It prevents the voice agent from saying something
• 14:51
  completely off -brand. Right. It brings necessary
• 14:53
  guardrails to enterprise customer service deployments.
• 14:56
  It solves the core trust issue holding voice
• 14:58
  AI back. Companies need absolute certainty before
• 15:01
  they deploy autonomous speaking agents. For sure.
• 15:04
  We've covered an immense amount of ground today.
• 15:07
  I want to pull all of these threads together
• 15:09
  cohesively. The big idea here is that the AI
• 15:12
  industry is completely bifurcating. On one end,
• 15:15
  you have massive, heavily guarded infrastructure
• 15:17
  being built. You have super intelligence models
• 15:20
  locked away for cyber defense. Those are the
• 15:22
  $95 billion server farms we discussed. Those
• 15:25
  are the specific models writing security patches
• 15:27
  with perfect precision. Exactly. On the other
• 15:30
  end, consumer AI is becoming deeply monetized
• 15:34
  rapidly. We see ads infiltrating search and exploding
• 15:37
  API costs everywhere. But it's also becoming
• 15:40
  hyper -practical for everyday normal users. We
• 15:44
  clearly see that with job hunting tools and smart
• 15:46
  AR glasses. We see it with enterprise voice checkpoints
• 15:50
  like Rhyme's Prism tool. We're moving from the
• 15:53
  initial wow phase to the industrial phase. The
• 15:55
  initial magic trick of chatting with a computer
• 15:58
  is over. Yeah, it really is. Now it's all about
• 16:00
  complex accounting, security, and massive infrastructure
• 16:03
  scale. That brings us to a final thought for
• 16:06
  you to consider. As these frontier models get
• 16:09
  smarter, they also get exponentially more expensive.
• 16:13
  Will the best AI only belong to massive corporations?
• 16:16
  Or will open source developers find a way to
• 16:19
  democratize superintelligence? That's the million
• 16:21
  dollar question. It's a question that will define
• 16:24
  the next decade of technology. Thank you for
• 16:26
  taking this deep dive with us today. Always a
• 16:29
  pleasure to unpack these complex shifts with
• 16:31
  you all. Take care and keep exploring.