Latest / AI Fire Daily / 🎙️ EP 319: Nvidia Sending Chips to the Moon & Microsoft's MAI Models Replace OpenAI

Transcript

🎙️ EP 319: Nvidia Sending Chips to the Moon & Microsoft's MAI Models Replace OpenAI · AI Fire Daily

• 0:00
  Imagine for a second the utter crushing cold.
• 0:04
  Oh, yeah. I mean, the space is just incredibly
• 0:07
  brutal. Right. The absolute silence, you know,
• 0:10
  of a dark lunar night. And sitting right there
• 0:12
  in the lunar dust is a tiny silicon microchip.
• 0:16
  Just freezing out there. It's completely alone
• 0:18
  in the dark, just waiting for the sun. Welcome
• 0:21
  back to the Deep Dive. I'm really glad you're
• 0:23
  joining us today. We have a truly fascinating
• 0:26
  journey ahead of us today. We really do. We're
• 0:28
  starting our trip way out in outer space. We're
• 0:31
  looking at AI physically landing on the moon.
• 0:34
  And then we're coming right back down to Earth.
• 0:36
  Exactly, to dissect a massive multi -billion
• 0:38
  dollar hardware war. We'll explore how big tech
• 0:41
  is ruthlessly cutting their costs. And finally,
• 0:44
  we'll look at everyday autonomous agents. Because
• 0:46
  AI is quietly shifting away from just being chatbots.
• 0:50
  The scale of these changes is... Genuinely breathtaking.
• 0:54
  Oh, absolutely. We're seeing computing hardware
• 0:57
  pushed to absolute physical limits and simultaneously
• 1:00
  AI software is becoming completely invisible
• 1:03
  and embedded. It's fundamentally changing how
• 1:06
  we interact with the digital world. Let's start
• 1:09
  by looking at those intense physical limits.
• 1:11
  NVIDIA is officially taking its GPU computing
• 1:14
  dominance off planet. Right. They are literally
• 1:17
  sending AI processing chips to the lunar surface.
• 1:21
  Which is wild. It really is. They're actually
• 1:23
  powering the next generation of autonomous space
• 1:26
  rovers. A space robotic startup called Lunar
• 1:30
  Outpost is leading this. Yeah, they're using
• 1:32
  NVIDIA's Jetson computing platform for their
• 1:35
  newest rover. They're using that chip to run
• 1:37
  a LiDAR system. And for those wondering, LiDAR
• 1:40
  involves using spinning lasers. It essentially
• 1:42
  maps physical spaces in 3D. Right. If this actually
• 1:46
  works, it makes aerospace history. It'll be the
• 1:48
  first GPU ever deployed on the moon. And the
• 1:50
  engineering hurdles they're facing are just staggering.
• 1:53
  These chips have to survive extreme cosmic radiation
• 1:56
  up there. Because that radiation can literally
• 1:59
  flip binary bits inside computer memory. Exactly.
• 2:02
  They also endure massive, incredibly violent
• 2:05
  temperature swings. And they must operate on
• 2:07
  incredibly low power. Because the lunar night
• 2:09
  is brutally cold and lasts for weeks. Yeah, it
• 2:12
  really does. Whoa, imagine processing complex
• 2:15
  mapping data while surviving the freezing lunar
• 2:18
  night. It feels almost like science fiction to
• 2:21
  me. It totally sounds wild, but it's actually
• 2:23
  highly practical. This specific mission is part
• 2:26
  of a much larger commercial push. Right. NASA
• 2:29
  wants to thoroughly map the entire lunar terrain.
• 2:32
  They're actively hunting for crucial resources
• 2:34
  like frozen water. Because this is all essential
• 2:37
  prep work for human astronauts, they want to
• 2:40
  send people back to the moon by 2028. And that
• 2:42
  timeline is exactly what's driving this rapid
• 2:45
  innovation. But NVIDIA isn't just stopping at
• 2:48
  surface rovers, right? No, not at all. They're
• 2:50
  also partnering with a company called Firefly
• 2:52
  Aerospace. They actually plan to put Jetson chips
• 2:56
  on a lunar satellite. This part is actually really
• 2:58
  interesting to me. The satellite will process
• 3:00
  mapping imagery directly in space. Yeah. It completely
• 3:03
  avoids beaming heavy raw data back to Earth.
• 3:06
  Yeah, that concept is known in the industry as
• 3:09
  edge computing. It means processing data locally
• 3:11
  instead of sending it away. So you don't send
• 3:14
  raw files to a distant Earth server. Right, because
• 3:16
  space communication inherently has terrible lag
• 3:19
  and very limited bandwidth. So it's basically
• 3:22
  like comparing a rugged off -grid survivalist
• 3:25
  to a massive factory. Oh, that's a great way
• 3:27
  to look at it. The jets and chips are those tiny,
• 3:30
  energy -efficient survivalists. Traditional data
• 3:33
  center GPUs are the massive, power -hungry factories.
• 3:37
  That analogy perfectly captures the extreme hardware
• 3:40
  constraints. And down the line, NASA is planning
• 3:43
  something even bigger. A massive astronaut carrying
• 3:46
  rover called Pegasus is currently in development.
• 3:49
  Though that specific timeline depends heavily
• 3:51
  on Blue Origin right now. They're still finishing
• 3:54
  crucial repairs on their heavy launch vehicle.
• 3:56
  So why is processing data locally in space so
• 3:59
  crucial? Well, transmitting raw, high -resolution
• 4:02
  mapping data across space takes forever. It ties
• 4:05
  up our deep space networks and wastes precious
• 4:07
  time. Local processing means you only beam back
• 4:10
  the final answers. So processing in space prevents
• 4:13
  a massive Earth -bound data bottleneck. Exactly.
• 4:16
  It fundamentally changes how efficiently we can
• 4:19
  explore the solar system. Processing data locally
• 4:21
  is clearly all about hardware efficiency. Back
• 4:25
  on Earth, this fight for hardware dominance is
• 4:28
  escalating rapidly oh it really is it's hitting
• 4:30
  the multi -billion dollar scale right now it
• 4:33
  really has become a massive high -stakes arms
• 4:36
  race yeah and amd is stepping up to the plate
• 4:38
  in a huge way they're investing up to five billion
• 4:40
  dollars into anthropic which is just a staggering
• 4:44
  amount of capital for one partnership right and
• 4:46
  starting in 2027 amd is supplying massive hardware
• 4:50
  They'll provide two gigawatts of MI450 GPUs.
• 4:53
  And, you know, GPUs are specialized chips handling
• 4:56
  thousands of AI calculations at once. Two gigawatts
• 4:59
  is a truly unbelievable amount of raw energy.
• 5:02
  To put that into perspective, a single gigawatt
• 5:05
  powers a whole city. Yeah, they're building computing
• 5:07
  infrastructure on a planetary scale now. It's
• 5:10
  not just about the chips. It's about power grids.
• 5:12
  The strategic goal here is twofold for them,
• 5:14
  right? Exactly. They want to expand Claude's
• 5:17
  computing power significantly, and they want
• 5:19
  to strengthen AMD's AI infrastructure business
• 5:22
  overall. So they're directly challenging their
• 5:24
  massive competitors in the space. Right. AMD
• 5:27
  desperately wants to prove they can handle the
• 5:29
  biggest workloads. Meanwhile, Anthropic needs
• 5:32
  reliable, massive compute to train smarter models.
• 5:36
  But I have to push back a little bit here. Okay,
• 5:38
  sure. Is two gigawatts of power purely about
• 5:41
  expanding computing capability? Or is this just
• 5:45
  a massive, necessary survival play? That's a
• 5:48
  very fair question. Because they need to ensure
• 5:51
  they aren't entirely dependent on NVIDIA. Yeah,
• 5:53
  survival is definitely the primary driver behind
• 5:55
  this deal. NVIDIA currently owns the vast majority
• 5:58
  of the AI hardware market. Right, and if you
• 6:00
  rely only on NVIDIA, you risk... major supply
• 6:03
  bottlenecks. Exactly. Anthropic needs guaranteed
• 6:06
  access to compute power to survive. And AMD needs
• 6:10
  a flagship partner to prove their hardware works.
• 6:12
  So it's a mutually beneficial defense strategy
• 6:14
  against a monopoly. That's exactly what it is.
• 6:17
  So what does this mean for the broader AI ecosystem?
• 6:19
  Well, we might finally see real sustained hardware
• 6:22
  competition. When multiple vendors compete, it
• 6:25
  drives down prices and increases innovation.
• 6:28
  Right. It proves Anthropic won't just rely on
• 6:30
  NVIDIA to scale up. It shows the entire industry
• 6:33
  that viable supply alternatives finally exist.
• 6:36
  Massive computing capability obviously requires
• 6:39
  massive ongoing financial costs, which is exactly
• 6:43
  why major tech players are changing their strategies.
• 6:46
  Yeah, they're starting to cut out the middlemen
• 6:48
  completely. Right. They're building highly capable
• 6:50
  models entirely in -house. Microsoft is probably
• 6:53
  the perfect example of this strategic shift.
• 6:55
  They just introduced two brand new in -house
• 6:59
  AI model variants. find them in public preview
• 7:02
  right now on Microsoft Foundry. They're giving
• 7:04
  enterprise users much more specific choices now.
• 7:07
  Exactly. You can balance image quality, voice
• 7:09
  speed, and daily operating costs. The first new
• 7:13
  model is called MAI Image 2 .5 Pro. And this
• 7:16
  is Microsoft's highest fidelity image generation
• 7:18
  model so far, right? It is. It generates beautiful,
• 7:21
  realistic images directly from simple text prompts,
• 7:24
  but it also edits your existing JPEG or PNG images
• 7:28
  seamlessly. So you just use natural language
• 7:30
  instructions. to change the picture. Exactly.
• 7:32
  And Microsoft is quietly replacing OpenAI's models
• 7:34
  across their entire ecosystem. They're putting
• 7:37
  their own custom models into Bing and PowerPoint.
• 7:40
  They're integrating them deeply into OneDrive,
• 7:43
  Dynamics 365, and Azure. The cost savings they're
• 7:47
  seeing internally are absolutely staggering.
• 7:50
  PowerPoint replaced GPT -Image 2 with this brand
• 7:53
  new Microsoft model. And they reduced the required
• 7:55
  GPU costs by up to 84%. Which is an incredible
• 7:59
  reduction in daily operational overhead. And
• 8:02
  they released the exact API pricing details for
• 8:05
  the public. It's $5 per 1 million text input
• 8:09
  tokens. And tokens are tiny word fragments that
• 8:13
  AI reads to process text. The image processing
• 8:15
  costs. are also highly competitive now. It's
• 8:17
  $8 for a million image input tokens. Right. And
• 8:20
  it costs $106 for image output tokens. They also
• 8:23
  released a new voice model called MAI Voice to
• 8:26
  Flash. This is the significantly faster and cheaper
• 8:28
  audio version, right? Yeah, it's a text to speech
• 8:30
  model built specifically for real time digital
• 8:33
  agents. It's perfect for automated call centers
• 8:36
  and high volume voice products. It offers very
• 8:38
  low latency, which means almost zero noticeable
• 8:41
  delay. So it's basically like a massive corporation
• 8:43
  going bankrupt. paying retail electricity rates.
• 8:48
  They finally just decided to build a nuclear
• 8:50
  reactor in their parking lot. I love that. They
• 8:52
  realized they had to generate their own power.
• 8:54
  That's exactly the financial dynamic driving
• 8:56
  this entire industry trend. MAI Image targets
• 9:00
  users who care deeply about precise visual quality.
• 9:03
  And MAI Voice targets enterprise products where
• 9:07
  sheer speed and cost matter most. You nailed
• 9:10
  it. Does this mean we're prioritizing speed and
• 9:13
  cost overall? Are we giving up having the absolute
• 9:16
  smartest model available? For everyday enterprise
• 9:19
  tasks, that is absolutely the tradeoff. You really
• 9:22
  don't need a genius level model to format a slide.
• 9:25
  You just need a model that is reliable. fast
• 9:28
  and cheap. The math simply doesn't support using
• 9:30
  massive models for mundane tasks. Yeah, good
• 9:34
  enough and cheap beats perfect and expensive
• 9:35
  for everyday tools. It really does. Sponsor placeholder.
• 9:38
  The ruthless cost cutting at the enterprise level
• 9:40
  is genuinely fascinating. It is because it actually
• 9:43
  mirrors shifting loyalties at the consumer level
• 9:45
  too. Oh, for sure. The walled gardens of these
• 9:48
  big platforms are finally cracking open. We're
• 9:51
  seeing a massive, undeniable shift in consumer
• 9:54
  user behavior. Google's Gemini platform has quietly
• 9:57
  reached 950 million monthly users. That's just
• 10:01
  a staggering number of active people. It's huge.
• 10:04
  Their user base has literally tripled since this
• 10:07
  time last year. Meanwhile... ChatGPT's overall
• 10:10
  market share just slipped below 50%. And a huge
• 10:14
  part of that is due to OpenAI's chaotic limits.
• 10:17
  Yeah, they keep constantly changing the rules
• 10:19
  for their premium subscribers. They've aggressively
• 10:21
  reset their usage limits 36 times recently. Changes
• 10:25
  so often, someone actually had to build a tracker
• 10:27
  site. I still wrestle with keeping up with OpenAI's
• 10:30
  changing limits myself. It's so annoying. It
• 10:32
  completely breaks your creative workflow when
• 10:34
  you hit a sudden wall. And rival companies are
• 10:36
  absolutely taking advantage of this user frustration.
• 10:38
  Oh, big time. Grok is making some very aggressive
• 10:41
  consumer moves right now. The new Grok build
• 10:44
  feature is officially heading to iOS soon. It
• 10:47
  lets you build custom apps, games, or data dashboards
• 10:49
  quickly. Exactly. You can instantly deploy them
• 10:52
  and publish on custom domains, and you can do
• 10:54
  it all without ever leaving the chat. They also
• 10:56
  just released their updated Grok 4 .5 model.
• 10:59
  Elon Musk openly admits it isn't quite as good
• 11:02
  as stable. Right. But he claims it's incredibly
• 11:04
  fast and very cost effective. It gets the daily
• 11:07
  job done for most average users. Musk claims
• 11:10
  Grok is cost effective for daily tasks. But when
• 11:13
  you look at raw performance benchmarks, the real
• 11:16
  dogfight Yeah, someone just tested Opus 5 .0
• 11:20
  directly against Fable 5 Max. Opus 5 .0 is currently
• 11:24
  in private testing right now. A creator posted
• 11:26
  this incredible full 3D comparison between them.
• 11:29
  Both models successfully completed the exact
• 11:32
  same complex task in one shot. It clearly shows
• 11:36
  how intensely close the flagship competition
• 11:38
  has actually become. We also have to mention
• 11:40
  Meta's recent marketing efforts. They released
• 11:42
  a very awkward new AI advertisement recently.
• 11:45
  That specific ad campaign was a very strange
• 11:47
  creative choice. The overarching slogan was simply,
• 11:50
  the future is for everyone. But they played a
• 11:52
  background song track about the earth dying.
• 11:55
  Yeah, it was a very awkward, dystopian contrast
• 11:57
  from Mark Zuckerberg's team. So are unpredictable
• 12:00
  platforms actively driving users away from their
• 12:03
  ecosystems? Definitely. Consumers have zero patience
• 12:06
  for constantly shifting digital boundaries. If
• 12:09
  a premium tool limits you randomly, you just
• 12:12
  find another tool. Consistency is usually much
• 12:15
  more important than being marginally smarter.
• 12:17
  Exactly. Unpredictable limits are driving frustrated
• 12:19
  users straight to Gemini. People just want reliable
• 12:22
  software that works when they actually need it.
• 12:25
  While platform loyalty shifts rapidly, the AI
• 12:28
  itself is rapidly evolving. It's moving away
• 12:31
  from just being a simple chatbox interface. Right.
• 12:34
  It's becoming an embedded autonomous agent in
• 12:36
  our daily lives. AI is starting to quietly manage
• 12:40
  our complex daily digital routines. This is arguably
• 12:43
  the most impactful technological trend we're
• 12:45
  seeing today. ChatGPT Health is now officially
• 12:48
  rolling out to all US users. It's currently available
• 12:50
  across every single active subscription plan.
• 12:53
  It natively integrates your personal medical
• 12:55
  records and daily wellness data. It can analyze
• 12:58
  your long -term health trends and give you insights.
• 13:01
  It's an incredibly powerful concept for tracking
• 13:03
  your personal health journey. You can spot anomalies
• 13:07
  in your blood work almost instantly. But I have
• 13:10
  some really deep reservations about this specific
• 13:12
  feature. Trusting AI models with private medical
• 13:16
  records feels incredibly risky right now. I completely
• 13:19
  understand that. Because these complex models
• 13:21
  can still hallucinate and just confidently make
• 13:24
  things up. You're completely right to be extremely
• 13:27
  cautious about this. Trusting it blindly with
• 13:30
  diagnostic medical decisions could get risky
• 13:32
  fast. You absolutely always have to verify what
• 13:35
  the language model tells you. They're also heavily
• 13:38
  expanding their standard desktop software capabilities.
• 13:41
  ChatGPT Voice on the desktop app just got a major
• 13:44
  update. Yeah, you can now operate multiple distinct
• 13:47
  agents across different digital workspaces. You
• 13:50
  can use ChatGPT Work and Codex at the exact same
• 13:52
  time. So the voice assistant manages everything
• 13:55
  seamlessly in the background for you. Exactly.
• 13:57
  We're also seeing entirely new tools built just
• 13:59
  for shared workflows. There's a new platform
• 14:01
  called Offloop that does this exceptionally well.
• 14:03
  Offloop gives specialized AI agents a shared
• 14:06
  digital workspace environment. Right. The autonomous
• 14:09
  agents actually collaborate on recurring business
• 14:12
  tasks together. It assigns the right specialized
• 14:15
  agent while keeping operational compute costs
• 14:17
  low. Visual and audio tools are also automating
• 14:21
  complex digital workflows now. Runway Media Router
• 14:24
  is a really great example of this trend. It automatically
• 14:27
  picks the absolute best AI model for your specific
• 14:30
  media. It actively balances image or video quality,
• 14:34
  generation speed, and overall cost. So you don't
• 14:37
  ever have to guess which model is currently best.
• 14:39
  Right. And Sunilo Sound Effects 1 .0 is doing
• 14:42
  this exactly for audio. Yeah. It generates perfectly
• 14:44
  synced sound effects directly from videos or
• 14:47
  text. It makes a clonal creation much faster
• 14:49
  and far more realistic. There's also this unreal
• 14:51
  GitHub 3D globe project floating around online.
• 14:55
  It's completely free and tracks massive amounts
• 14:57
  of unstructured global data. It actively tracks
• 14:59
  over 500 live global news feeds simultaneously.
• 15:03
  Yeah, it maps active wars, flights, financial
• 15:05
  markets, shipping, and cyber incidents. It synthesizes
• 15:08
  all that global chaos onto one live 3D globe
• 15:11
  interface. It's the perfect visual example of
• 15:15
  AI aggregating complex real -time data. It makes
• 15:20
  understanding our chaotic world much easier to
• 15:23
  visually digest. So what happens when AI agents
• 15:26
  start managing other AI agents? Well, we fundamentally
• 15:29
  step out of the loop of daily execution. We become
• 15:32
  strategic directors instead of being hands -on
• 15:35
  operators. We just give the high -level operational
• 15:37
  goals to the overall system. We're moving from
• 15:40
  prompting AI to simply managing a digital workforce.
• 15:43
  The AI completely handles the logistics and we
• 15:46
  just handle the strategy. We've covered a tremendous
• 15:48
  amount of intellectual ground today. Let's synthesize
• 15:50
  this entire technological journey for you. Yeah,
• 15:53
  we started with extreme hardware survival on
• 15:55
  the freezing lunar surface. We explored massive
• 15:58
  billion dollar compute deals back here on Earth.
• 16:00
  We clearly saw how major tech companies are ruthlessly
• 16:03
  cutting their costs. Right. They're heavily favoring
• 16:06
  operational efficiency and speed over absolute
• 16:08
  raw intelligence. All of this development leads
• 16:10
  to a very specific new world. AI is quietly becoming
• 16:13
  deeply embedded into absolutely everything we
• 16:16
  use. From mundane PowerPoint presentations to
• 16:20
  our highly personal private medical records.
• 16:23
  It's rapidly becoming a completely invisible
• 16:25
  layer of our daily lives. Just like background
• 16:28
  electricity, we'll very soon stop noticing it
• 16:31
  entirely. I want to leave you with a final thought
• 16:33
  today. Oh, this is the best part. If AI is becoming
• 16:36
  an invisible, deeply embedded utility everywhere.
• 16:40
  Much like electricity powering a lunar satellite
• 16:43
  or your office desktop. What happens to our own
• 16:46
  critical thinking over time? That's a big one.
• 16:48
  What happens when we no longer even notice AI
• 16:50
  making decisions for us? That is easily the most
• 16:53
  important question we can ask right now. Thank
• 16:55
  you so much for taking this deep dive with us
• 16:56
  today. Keep staying intensely curious and keep
• 16:59
  questioning the tech you use. We'll see you next
• 17:01
  time.