Latest / Today’s AI News / OpenAI’s AI Phone Is Coming in 2027, Mini Data Centers for Your Home, Anthropic Goes Finance

Transcript

OpenAI’s AI Phone Is Coming in 2027, Mini Data Centers for Your Home, Anthropic Goes Finance · Today’s AI News

• 0:00
  Welcome back to the day's AI News, brought to you by 9X
• 0:02
  Productions. Hey everyone, great to be here
• 0:04
  for this one. Yeah, we have a crazy lineup
• 0:06
  today, so I want you to imagine your smartphone isn't just, you
• 0:11
  know, a screen you tap, but a digital Butler that actually
• 0:14
  physically sees the world through a camera lens and acts
• 0:18
  on your behalf. Right, like a real assistant.
• 0:20
  Exactly. We've been dreaming of that leap
• 0:23
  from a digital interface to a real world assistant, and the
• 0:27
  timeline just got completely upended.
• 0:29
  Open AI is officially trying to put that Butler in your pocket
• 0:32
  much, much sooner than we thought.
• 0:34
  Which is just wild to think about.
• 0:35
  It is O Our mission for this Dee dog today is to explore how AI
• 0:40
  is making this massive leap from just, you know, software on a
• 0:44
  screen into our physical world. Right from the hardware in your
• 0:47
  pocket to data centers strapped to your house.
• 0:49
  Yeah, literally strapped to your house.
• 0:51
  Yeah, and even hyper specialized agents doing Wall Street's heavy
• 0:54
  lifting. We've got supply chain reports,
• 0:56
  technical white papers, joint venture filings, all of it.
• 0:59
  We are basically watching artificial intelligence escape
• 1:02
  the browser window. Like for the last few years AI
• 1:06
  has been a destination right? A website you type into.
• 1:09
  Yeah, chadgt.com. Exactly.
• 1:12
  But now it's being woven directly into the physical and
• 1:15
  economic infrastructure around us.
• 1:18
  O let's start with that digital Butler thing because the
• 1:20
  timeline here is, well, it's mind boggling.
• 1:23
  We just got a new report from supply chain analyst Ming Chi
• 1:26
  Ku. Oh, Kuo rarely misses.
• 1:29
  Right. If you follow tech, you know his
• 1:30
  track record is incredibly solid.
• 1:33
  So he's saying Open AI is fast tracking its very first AI agent
• 1:36
  phone, aiming for mass production in the first half of
• 1:39
  2027. Wow. 2027.
• 1:42
  Yeah, a full year earlier than anyone expected.
• 1:45
  I mean, a year early in the hardware world is practically
• 1:47
  lightspeed. Hardware supply chains are
• 1:49
  usually locked in, you know, years in advance.
• 1:51
  Tooling, factories and all that. Exactly.
• 1:53
  Tooling, securing raw materials, booking fabrication time at
• 1:56
  TSMC. It takes massive foresight.
• 1:58
  So this aggressive timeline tells me Open AI feels immense
• 2:01
  pressure to lock down a physical device before the market
• 2:04
  settles. And the specs are what really
• 2:06
  caught my eye. O Mediatek is apparently the
• 2:09
  sole chip supplier, and they're using two distinct AI Rocessors.
• 2:13
  Oh, interesting. Not just one central brain.
• 2:15
  Nope. The idea is to handle language
• 2:18
  and vision tasks simultaneously. It's kind of like the brains
• 2:21
  hemispheres, right? One part is looking at the
• 2:23
  environment, the other is talking and reasoning about it.
• 2:26
  That makes. Total sense for an agent.
• 2:27
  Yeah, and there's a huge focus on the image signal processor.
• 2:30
  They're using an enhanced HDR high dynamic range pipeline just
• 2:34
  for real world visual sensing, right?
• 2:37
  And Kuo estimates they could ship 30 million units between
• 2:41
  2027 and 2028. Word is this rush is driven by
• 2:45
  open AI wanting to IPO Plus just the insane competition right
• 2:50
  now. Well, the dual processor
• 2:51
  architecture tells us a lot about how this thing will
• 2:53
  actually work. Because, you know, in a normal
• 2:55
  smartphone, a single processor juggles everything.
• 2:57
  Right, it switches back and forth.
• 2:59
  Exactly. And AI models require massive
• 3:01
  matrix multiplications if your phone is using its neural engine
• 3:05
  to process live video of your kitchen and you ask it a
• 3:08
  question. It creates lag.
• 3:10
  Right. It has to context switch, which
• 3:11
  creates latency. But by dedicating one piece of
• 3:15
  silicon entirely to vision and another to language, you
• 3:18
  basically eliminate that lag. The agent can talk to you and
• 3:21
  watch a live feed without skipping a bite.
• 3:23
  And that ties right into the HDR focus.
• 3:26
  HDR takes multiple exposures at different light levels and
• 3:29
  merges them. Because if an AI is going to
• 3:31
  look under your sink to fix a pipe or I don't know, inside a
• 3:35
  dark fridge for a recipe. It needs perfect vision.
• 3:38
  Exactly. Real life isn't a photography
• 3:40
  studio. HDR gives the AI a visual cortex
• 3:45
  built for messy, uneven lighting.
• 3:47
  Right, because bad visual data just leads to bad AI reasoning.
• 3:51
  But taking a step back, the hardware strategy itself is a
• 3:54
  huge pivot because controlling the hardware and the OS is
• 3:57
  really the only way to build a true agent phone.
• 4:00
  You mean instead it's just an app?
• 4:01
  Exactly. If Open AI relies on Apple or
• 4:04
  Google's OS, they're instantly bottlenecked.
• 4:07
  Apple could just restrict background camera access for
• 4:09
  privacy or limit battery usage. Or force payments through their
• 4:12
  App Store. Exactly.
• 4:14
  By owning the full stack, the AI can act on your behalf 24/7
• 4:18
  without asking a Silicon Valley middleman for permission.
• 4:22
  But wait, let me push back on this a bit.
• 4:24
  If this Mediatek phone drops in 2027, where does this leave that
• 4:28
  mysterious device they're building with Joni Ives IO
• 4:31
  brand? The ambient device, yeah.
• 4:34
  Are they the same thing, or is open AI kind of, you know,
• 4:37
  fracturing its hardware focus here?
• 4:39
  I mean, it's entirely possible they're running parallel tracks.
• 4:42
  The Joni I've project might be this experimental screenless
• 4:45
  wearable, while the Mediatek phone is the mass market play.
• 4:48
  Because people already understand a rectangle with a
• 4:50
  screen. Exactly.
• 4:51
  If you want 30 million users by 2028, you need a familiar form
• 4:55
  factor. But either way, they both need
• 4:57
  vertical integration. They need an environment where
• 5:00
  their models are the operating system.
• 5:02
  OK. So if they actually put 30
• 5:03
  million sensing AI phones in our pockets, the compute power
• 5:08
  needed for that is just mind boggling.
• 5:10
  It's massive. And instead of building giant
• 5:12
  warehouses, the solution might literally be strapped to the
• 5:15
  side of your house. Which is such a wild transition.
• 5:18
  The footprint of AI is moving from the cloud right into the
• 5:21
  suburbs. Literally.
• 5:22
  So there's this California startup called Spam, and they're
• 5:25
  teaming up with NVIDIA to install what they call mini AI
• 5:29
  data centers, Exfra units on the exterior walls of homes and
• 5:33
  small businesses. Just right there on the siding.
• 5:35
  Right on the wall they tap into the unused electrical capacity
• 5:38
  on local grids, and inside they're packing Nvidia's liquid
• 5:42
  cooled RTXPRO 6000 Blackwell GPU's, that is.
• 5:48
  Serious hardware. Seriously, Span says they can
• 5:51
  install 8000 of these units 6 times faster and for 1/5 of the
• 5:55
  cost of a traditional centralized 100 MW data center.
• 5:59
  And they're already testing this with Poulter Group on new
• 6:01
  residential builds. It makes sense when you look at
• 6:03
  the US power grid. The bottleneck right now isn't
• 6:06
  generating electricity, it's transmission.
• 6:08
  Getting the power to the massive data centers.
• 6:10
  Exactly. Tech giants have to wait like
• 6:12
  five to seven years for utilities to build high voltage
• 6:15
  lines. Span completely bypasses that.
• 6:17
  Neighborhoods already have power lines that sit underutilized
• 6:20
  during the day or late at night. So they're just soaking up the
• 6:23
  excess capacity. Right, distributing the compute
• 6:25
  across thousands of homes. OK, sure, it sells the grid
• 6:29
  strain on paper, but realistically, are parents
• 6:32
  really going to accept a humming AI server mounted right where
• 6:36
  their kids play in the backyard? That is the big question.
• 6:39
  I mean, we see fights over smart meters.
• 6:43
  A suburban data center feels like a tough sell, especially
• 6:46
  with ocean and space data centers also being developed.
• 6:48
  The community pushback is going to be huge.
• 6:51
  People naturally distrust black box tech on their property, but
• 6:55
  that's why the liquid cooling is the key.
• 6:58
  Because it's quiet. Exactly, Blackwell GPU's
• 7:00
  generate immense heat. In a normal data center you have
• 7:03
  rows of screaming server fans. You obviously can't put that on
• 7:06
  a house. Liquid cooling absorbs the heat
• 7:09
  almost silently, like a car radiator.
• 7:11
  Oh I see, and if it's silent and span subsidizes your energy bill
• 7:14
  or pays a lease fee for the wall space?
• 7:16
  Economics might win out. Right.
• 7:18
  The industry is so desperate for power that socially risky ideas
• 7:22
  like residential data boxes are becoming economically viable.
• 7:25
  Wow, so while Compute is moving into the suburbs, the AI models
• 7:29
  themselves are moving straight to Wall Street.
• 7:31
  Yes, specializing heavily. Right.
• 7:34
  Anthropic is taking a totally different approach to agents
• 7:36
  than open AI's consumer phone. They just unveiled 10 Ready to
• 7:40
  Run AI agents built specifically for financial services and
• 7:44
  insurance. Leaving the consumer stuff
• 7:46
  behind for the enterprise. Exactly, and these aren't just
• 7:49
  chat bots, they handle things like building pitch books,
• 7:52
  screening, know your customer KYC files, reviewing earnings.
• 7:56
  Real analyst work. Yeah, and they come with domain
• 7:59
  specific instructions, add on clawed models and data
• 8:01
  connectors from partners like Dunn and Bradstreet, Verisk and
• 8:05
  Ibis World. The data connectors are the game
• 8:07
  changer there. Totally.
• 8:09
  Firms can customize them to their own risk policies and run
• 8:12
  them via Claude Co work, Claude code or manage agents.
• 8:16
  Plus, Anthropic announced a $1.5 billion joint venture with Wall
• 8:20
  Street Giants for this. Wow.
• 8:22
  It feels like a huge divide. Like Open AI is trying to sell
• 8:25
  us the brakes to build our own consumer tools, but Anthropic is
• 8:28
  selling Wall Street the fully furnished house.
• 8:30
  Or an armored truck, really. An armored truck exactly with
• 8:33
  the GPS programmed and a compliance officer inside.
• 8:37
  Right, because when an analyst does a KYC check, they pull from
• 8:40
  dozens of databases for sanctions and risk.
• 8:43
  A standard LLM struggles to read that unstructured text and
• 8:46
  hallucinations are a huge risk. Which banks can't afford?
• 8:50
  Exactly, and Fropix agents use AP is to pull pristine
• 8:54
  structured data directly from places like virisk.
• 8:57
  It analyzes it against compliance rules without
• 8:59
  guessing. It's just verifying facts.
• 9:02
  Yeah, in finance, accuracy is legally mandated.
• 9:06
  Anthropic is meeting highly regulated industries exactly
• 9:09
  where they are, rather than selling a generalized model and
• 9:12
  hoping they figure it out. Makes total sense, but hey, you
• 9:15
  don't need a $1.5 billion Wall Street budget to get autonomous
• 9:18
  agents working for you right now.
• 9:20
  You can literally build one in Notion today, and this requires
• 9:23
  a shift from chatting with a bot to asynchronous computing.
• 9:27
  This is that enterprise trickling down to the consumer
• 9:29
  level, yes. Yeah, there's a hidden workflow
• 9:32
  to make Notion agents autonomous.
• 9:33
  Here's how you do it. First create an agent and
• 9:36
  reports database. OK, I got the database.
• 9:38
  Second, create a new template and name it with the AT today
• 9:41
  tag like at today daily summary. That way it inherits the current
• 9:45
  date Glover. Third, write your instructions
• 9:48
  in the page body and AT mention the agent.
• 9:51
  But, and this is crucial, make sure to stop the agent while
• 9:55
  you're building this. Otherwise it triggers
• 9:57
  immediately. Exactly, and it overwrites your
• 9:59
  template. Finally, duplicate the template
• 10:01
  on a specific cadence like daily at 7:00 AM.
• 10:05
  What you're doing there is basically mimicking A Cron job.
• 10:08
  A Cron job. Yeah, in computer science, it's
• 10:10
  a utility that runs scripts at scheduled times.
• 10:12
  By setting that Notion template on a daily cadence, The
• 10:15
  duplication triggers the agent. You're moving to asynchronous
• 10:19
  AI. The agent wakes up, does the
• 10:21
  work, and goes back to sleep. Pro tip for everyone listening
• 10:25
  Use this exact workflow to route your daily debris and weekly
• 10:28
  automated emails through a single planning agent.
• 10:32
  You're effectively giving yourself a personal chief of
• 10:34
  staff. It changes your whole
• 10:35
  relationship with the software. You're a manager of a system
• 10:37
  now, not just an operator. Amazing.
• 10:40
  OK, this shift toward agenic AI is happening everywhere, all at
• 10:43
  once. Let's hit the rest of today's
• 10:44
  landscape in a rapid fire rundown.
• 10:46
  Let's do it starting with tools. Pomelli is making product
• 10:49
  specific campaigns with Google's AI marketing tool and Luma
• 10:53
  dropped Uni 1.1, a new image generation and editing AI that
• 10:58
  is nearing frontier level. The visual quality on that is
• 11:01
  getting incredible. Yeah, it really is.
• 11:03
  On the news side, Open AI is rolling out GT 5.5 Instant to
• 11:07
  all users. It's got stronger memory and
• 11:10
  more concise responses which. Is always nice.
• 11:12
  Definitely. Microsoft expanded its Copilot
• 11:14
  Co Work Agenix system to iOS and Android.
• 11:17
  Which makes sense. If agents are asynchronous, they
• 11:19
  need to live on our phones, not just our laptops.
• 11:22
  Exactly. Meanwhile, Apple is paying $250
• 11:25
  million to settle a class action suit over misleading claims
• 11:28
  about its new AI. Siri Coach.
• 11:30
  Yeah, over in finance perplexity.
• 11:32
  AI Launch Computer for Professional Finance with 35
• 11:35
  dedicated workflows and Tropic committed to spending $200
• 11:38
  billion on Google's cloud and chips over five years. 200
• 11:42
  billion. That is a massive scale.
• 11:44
  Huge. And lastly, Coinbase is cutting
• 11:48
  14% of its workforce, about 700 people to shift toward AI,
• 11:52
  native teams and agent driven workflows.
• 11:55
  That is a stark indicator for enterprise efficiency.
• 11:57
  It's a wild juxtaposition, right?
• 12:00
  Apple's paying 1/4 billion dollars because they over
• 12:02
  promised what AI could do, while at the exact same time Coinbase
• 12:06
  is laying off 14% of its staff because AI is doing so much.
• 12:09
  It's a paradox, but it perfectly illustrates the maturity curve
• 12:13
  right now. How so?
• 12:14
  There's a rule of thumb emerging.
• 12:16
  If a task needs general context and common sense, AI is a
• 12:20
  liability. But if it requires rigid
• 12:22
  structure, AI is a job killer. Oh, that makes sense.
• 12:25
  Sirius, General. Right, Siri operates in the
• 12:28
  messy, unstructured world of consumer requests.
• 12:30
  It's struggles, but Coinbase operates in structured code and
• 12:34
  financial workflows. In that rigid context, AI wildly
• 12:38
  over delivers. That is fascinating.
• 12:40
  Well, let's step away from the massive corporate shakeups and
• 12:42
  look at how actual people like you listening right now are
• 12:44
  using this tech in the real world.
• 12:46
  I love these community workflows.
• 12:47
  Me too. Today we have one from a reader
• 12:50
  named Shy West in Springfield, IL.
• 12:54
  So Shy went to a multi day music festival.
• 12:56
  Sounds chaotic. Totally chaotic, overlapping set
• 13:00
  times, Multiple locations. So they fed Claude the
• 13:03
  festival's schedule PDF. They told it to ignore all the
• 13:06
  workshops and autograph signings and gave it a list of must see
• 13:10
  bands versus nice to see bands. Oh, that is brilliant.
• 13:13
  Right. Claude outputted a perfectly
• 13:15
  optimized spreadsheet schedule, noting partial sets, stage
• 13:19
  times, and even its decision making rationale.
• 13:22
  They basically ask Claude to solve the traveling salesman
• 13:25
  problem for a music festival. Exactly, and it nailed it.
• 13:29
  It's the perfect use case for a human.
• 13:31
  Holding all those overlapping variables in your working memory
• 13:33
  is super stressful. Oh, I'd be highlighting things
• 13:35
  for hours. Exactly.
• 13:37
  But an AI model can hold thousands of permutations in its
• 13:40
  context window instantly. It takes a complex multi
• 13:43
  variable logistical problem and just solves it so you can
• 13:46
  actually enjoy the event. Such a great example.
• 13:49
  You know, if you look at everything we've talked about
• 13:51
  today, from open an AI putting an agent in your pocket to span
• 13:55
  bolting a data center to your house, the line between
• 13:59
  technology we use and the physical infrastructure we live
• 14:02
  inside is just completely disappearing.
• 14:04
  It really is. It leaves you with this
• 14:05
  question, how long until our entire physical environment is
• 14:09
  just one giant interconnected agent processing data on our?
• 14:13
  Behalf, that is a thought to chew on.
• 14:16
  Well, make sure to subscribe so you stay updated on all of this.
• 14:19
  And if you're enjoying the show, we'd love it if you took a
• 14:22
  second to rate US five stars. We will see you back here for
• 14:25
  tomorrow's deep dive.