Latest / Connecting the Dots / OpenAI's Stargate Pause, Anthropic's Legal Battle, and Meta's Superintelligence Push

Transcript

OpenAI's Stargate Pause, Anthropic's Legal Battle, and Meta's Superintelligence Push · Connecting the Dots

• 0:00
  Today's tech news moves fast, but understanding
• 0:03
  it shouldn't be hard. That's why Snarfl Solutions
• 0:06
  Group brings you Connecting the Dots, where we
• 0:09
  break down what's happening in tech and why it
• 0:11
  matters to you. Now, let's dive in with your
• 0:14
  hosts. Welcome to Connecting the Dots from Snarfl
• 0:18
  Solutions Group, right here in Sacramento, California.
• 0:20
  I'm Alex. And I'm Morgan. It is Thursday, April
• 0:23
  9th. 2026 and uh if you're sitting in sacramento
• 0:27
  with us right now it is just a gorgeous mostly
• 0:30
  sunny day we're heading toward a high of around
• 0:32
  76 degrees yeah perfect spring weather exactly
• 0:35
  and the financial markets are also looking pretty
• 0:37
  sunny i mean the dow jumped over 931 points yesterday
• 0:40
  which is huge right it closed near 48 000 the
• 0:44
  s &p is up 28 and bitcoin is just uh hovering
• 0:48
  comfortably around 71 000 It definitely feels
• 0:51
  like a very normal, prosperous day. I will say,
• 0:54
  though, if you are traveling, keep an eye out.
• 0:56
  The National Weather Service is actually managing
• 0:58
  254 active weather alerts across the country.
• 1:01
  Oh, wow. Yeah, definitely check the radar. But
• 1:04
  I mean, mostly it's a day defined by human routines,
• 1:07
  right? Sports, stable markets. If you're a baseball
• 1:09
  fan, last night was fantastic. The Giants secured
• 1:12
  a beautiful 5 -0 shot out against the Phillies.
• 1:14
  With Matt Gage on the mound, right? Yep. He was
• 1:16
  pitching. And the Athletics pulled off a really
• 1:19
  tight 3 -2 victory over the Yankees. So, a good
• 1:22
  night for baseball. Yeah! But, you know, while
• 1:25
  you're out there sipping your morning coffee
• 1:27
  and checking those baseball scores, There is
• 1:29
  this completely different, almost invisible reality
• 1:32
  running concurrently in the background. Oh, for
• 1:35
  sure. We are watching the skyrocketing trillion
• 1:38
  dollar ambitions of the artificial intelligence
• 1:41
  industry suddenly slam into very hard, very physical
• 1:46
  brick walls. Oh. And that is really what we want
• 1:49
  to focus on today. Yeah, let's unpack this because
• 1:51
  AI is no longer just this magical software running
• 1:54
  on a screen somewhere. Right. It is aggressively
• 1:56
  demanding massive amounts of physical space.
• 1:59
  It's draining actual electrical grids. And it's
• 2:02
  actively forcing sovereign governments in the
• 2:04
  military to react. We are literally witnessing
• 2:07
  the moment the digital cloud is forced to come
• 2:09
  down to the physical earth. Exactly. So we're
• 2:12
  going to break down the day's top tech stories
• 2:14
  from the Morningstar and see what this means
• 2:16
  for all of us. But first, today's episode is
• 2:18
  brought to you by Strong Coffee, the premium
• 2:21
  instant coffee that tastes like a traditional
• 2:23
  brew but is packed with 15 grams of protein,
• 2:26
  MCTs, and adaptogens like ashwagandha. Get clean,
• 2:30
  sustained energy without the jitters. Use promo
• 2:33
  code SNARFL for 20 % off at strongcoffeecompany
• 2:36
  .com. All right, so jumping into our first story,
• 2:38
  if you really want to understand how hard this
• 2:41
  collision between AI and reality is, you have
• 2:44
  to look at the sheer physics of electricity.
• 2:46
  Yeah, which brings us to OpenAI. They have officially
• 2:49
  hit the brakes on a massive overseas project
• 2:51
  called Stargate UK. Right. And for some context
• 2:54
  here, Stargate is this mind -bogglingly large
• 2:57
  $500 billion global infrastructure plan. Half
• 3:00
  a trillion dollars. Literally half a trillion.
• 3:03
  And the initial equity funders are heavyweights,
• 3:05
  right? You've got OpenAI, Oracle, MGX, SoftBank.
• 3:08
  They're trying to build the physical data centers
• 3:10
  that will train the next generation of AI. But
• 3:12
  a project that size is almost hard to comprehend.
• 3:15
  I mean, we aren't talking about putting a few
• 3:17
  servers in a closet somewhere. No, not at all.
• 3:19
  Stargate already has facilities operating in
• 3:21
  Texas. They have active builds right now in the
• 3:24
  UAE and Norway. Right. But the UK branch, which,
• 3:27
  by the way, was just announced last September
• 3:29
  alongside Nvidia and a London based company called
• 3:32
  Nescale, that entire branch is now being completely
• 3:36
  shelved. And the reasons OpenAI gave for pulling
• 3:39
  the plug are incredibly telling. They cited the
• 3:42
  high cost of energy and a burdensome regulatory
• 3:45
  environment. Yeah, they're essentially packing
• 3:46
  up and saying well, we're only going to move
• 3:48
  forward when the local conditions actually allow
• 3:50
  for it. But I guess the question for you is why
• 3:53
  is this happening? Like, why is energy suddenly
• 3:56
  the ultimate deal -breaker for OpenAI, a company
• 3:59
  that was recently valued at $852 billion? It
• 4:03
  really comes down to the mechanics of how AI
• 4:05
  actually learns and operates. Because, you know,
• 4:07
  when you do a standard Google search, you are
• 4:09
  essentially just asking a computer to fetch a
• 4:11
  file from a library. Which takes almost nothing.
• 4:14
  Exactly. Very little power. But when you ask
• 4:17
  an advanced AI a question, it isn't fetching
• 4:20
  a file. It is actively performing billions of
• 4:23
  complex parallel matrix multiplications. It's
• 4:26
  generating the answer from scratch. Right then
• 4:28
  and there. Right. Across... thousands of GPUs
• 4:31
  graphics processing units just to predict the
• 4:34
  perfect sequence of information and doing that
• 4:36
  requires an astonishing amount of electrical
• 4:38
  current and it doesn't stop there right because
• 4:40
  of the heat oh the heat is a massive issue when
• 4:43
  you push electricity through that many processors
• 4:45
  you generate insane amounts of heat so for every
• 4:48
  watt of power you use to run the actual computer
• 4:51
  you need an enormous amount of power and water
• 4:53
  just to run the cooling systems. Just so the
• 4:56
  building doesn't literally melt down. Pretty
• 4:58
  much. And Nenscale, the company partnering with
• 5:00
  OpenAI in the UK, they just raised $2 billion
• 5:03
  themselves. They were planning to scale up to
• 5:06
  31 ,000 NVIDIA GPUs in a single facility. I mean,
• 5:10
  just try to imagine plugging 31 ,000 industrial
• 5:13
  ovens into a single city's power grid and turning
• 5:16
  them all on at the exact same time. You'd blow
• 5:18
  the grid. You can't just flip the switch. You
• 5:20
  would completely brown out the entire region.
• 5:23
  So that is the physical reality. But there are
• 5:25
  massive political consequences here, too, especially
• 5:28
  for UK Prime Minister Keir Starmer. Yeah. Losing
• 5:32
  this project is a devastating blow for him. He
• 5:35
  has spent the last two years really pushing to
• 5:37
  turn the UK into an AI growth zone. Right, because
• 5:41
  this specific open AI data center was slated
• 5:44
  to bring in 5 ,000 jobs and 30 billion pounds
• 5:47
  in private investment. The political hype was
• 5:49
  huge. I remember they even touted it during Donald
• 5:51
  Trump's state visit as this massive cross -national
• 5:54
  victory. Which really frames the situation perfectly.
• 5:57
  I mean, think about it like you're trying to
• 5:59
  build a mega skyscraper. OK, yeah. You can have
• 6:01
  billions of dollars sitting in the bank. You
• 6:03
  can have the most brilliant blueprints drawn
• 6:06
  up by the best architects in the world. But if
• 6:08
  the local utility company literally cannot wire
• 6:11
  enough electricity to the site and the city won't
• 6:14
  approve your environmental permits, your billion
• 6:16
  dollar skyscraper is just going to stay an empty
• 6:18
  dirt lot. Exactly. And I think this exposes a
• 6:21
  really glaring flaw in the whole concept of sovereign
• 6:24
  AI. Oh, absolutely, because the big push lately
• 6:27
  is that every country needs its own AI, right?
• 6:31
  Running locally within its own borders to protect
• 6:34
  national data and legal jurisdiction. Right.
• 6:36
  But if a highly developed, wealthy nation like
• 6:39
  the UK can't provide the raw electrical grid
• 6:42
  stability and the regulatory speed required to
• 6:45
  house these machines, it really casts serious
• 6:48
  doubt on whether sovereign AI is anything more
• 6:50
  than a pipe dream for most of the world. Yeah,
• 6:52
  it's a huge wake up call. So if the physical
• 6:54
  power grid is the first massive bottleneck to
• 6:57
  the AI revolution, the second is sovereign control.
• 7:01
  Right. Moving on to our next story. Yeah, because
• 7:03
  once you actually find the electricity and build
• 7:05
  these massive physical machines, the immediate
• 7:08
  question becomes who gets to command them during
• 7:10
  a national crisis. And that tension is exploding
• 7:13
  right now. There is a massive legal battle going
• 7:15
  on between Anthropic and the Pentagon. This is
• 7:17
  such a wild story. It really is one of the most
• 7:20
  fascinating, ethical, and legal tugs of war we've
• 7:23
  seen in modern tech. Basically, a D .C. appeals
• 7:26
  court just denied Anthropic's request to temporarily
• 7:30
  block the Department of Defense from designating
• 7:32
  them as a supply chain risk. Which is a huge
• 7:35
  reversal because back in March, Anthropic actually
• 7:39
  won a preliminary injunction in California. Right,
• 7:41
  and the judge there actually criticized the Pentagon's
• 7:44
  move. Yeah. But this new ruling from the DC Circuit
• 7:46
  Court of Appeals stated that the equitable balance
• 7:49
  here cuts in favor of the government. So because
• 7:53
  of this official label, defense contractors are
• 7:55
  strictly banned from using Anthropic's Claude
• 7:58
  AI models for any DOD work. And to really understand
• 8:01
  the gravity of this, we have to look at the mechanics
• 8:03
  of the conflict. So Anthropic, led by CEO Dario
• 8:07
  M .O. Day, they built their entire brand around
• 8:10
  AI safety. Right. That's their whole thing. Yeah.
• 8:12
  They hard code strict ethical guardrails into
• 8:15
  their models to prevent the AI from assisting
• 8:17
  in violence or cyber attacks or weapons development.
• 8:20
  The DoD wanted to use Anthropix models, but they
• 8:23
  demanded that those safety restrictions be removed
• 8:25
  for military use. And Anthropix refused. So,
• 8:29
  just to be clear. We are completely neutral on
• 8:32
  this. We aren't taking sides. We're just laying
• 8:34
  out what the sources are saying about both arguments.
• 8:37
  Exactly. Just reporting the facts. So on one
• 8:39
  side, you have Anthropic arguing that the government
• 8:43
  labeling them a supply chain risk is arbitrary
• 8:46
  unconstitutional retaliation. They are essentially
• 8:49
  saying, you know, you are punishing us financially
• 8:51
  because we refuse to compromise our corporate
• 8:53
  ethics. And then on the other side, you have
• 8:56
  acting U .S. Attorney General Todd Blanch arguing
• 8:59
  for military readiness. Right. His legal argument
• 9:02
  is that operational control belongs solely to
• 9:05
  the commander in chief in the Department of Defense,
• 9:07
  not to a board of tech executives in Silicon
• 9:10
  Valley. Which makes sense from a defense standpoint.
• 9:12
  Right, because if an AI is going to be integrated
• 9:15
  into sensitive military logistics or targeting
• 9:17
  systems, the military feels they must have absolute
• 9:21
  unrestricted access to the underlying logic of
• 9:24
  that tool. It's like, imagine you're a pacifist
• 9:27
  tool manufacturer, right? And you secure a contract
• 9:29
  to supply Swiss Army knives to a combat battalion.
• 9:33
  But because of your corporate ethics, you refuse
• 9:36
  to include the blade. You only provide the screwdriver
• 9:39
  and the corkscrew. The military is telling you,
• 9:42
  look, we are in a high stakes combat environment.
• 9:45
  We cannot use a crippled tool. If you won't give
• 9:48
  us the unrestricted version, you are a liability
• 9:51
  to our supply chain and you're off the vendor
• 9:53
  list. That captures the operational friction
• 9:56
  perfectly. The DC court essentially had to weigh
• 9:58
  two things. The financial harm to a private tech
• 10:02
  company versus the judicial management of how
• 10:05
  the Department of Defense secures vital technology
• 10:07
  during active conflicts. And they decided that
• 10:10
  military security supersedes corporate safety
• 10:12
  protocols? Exactly. But wait. Looking at the
• 10:15
  financial numbers from our sources, the actual
• 10:18
  harm to Anthropic seems completely non -existent.
• 10:21
  No, they are doing fine. They are absolutely
• 10:23
  thriving commercially. I mean, their annual revenue
• 10:25
  just skyrocketed from $9 billion at the end of
• 10:29
  2025 to over $30 billion now. Yeah, it's a massive
• 10:32
  jump. And their valuation is sitting at $380
• 10:36
  billion. They're even rolling out specialized
• 10:39
  models like Claude Mythos for cybersecurity,
• 10:41
  and they're partnering with Google and Apple.
• 10:43
  Right. So I have to ask, do they actually care
• 10:46
  about losing this DoD contract financially? Or
• 10:49
  is this entire court battle just a massive PR
• 10:52
  flex over who gets to dictate the rules of AI?
• 10:56
  Honestly, the financials prove this is almost
• 10:59
  entirely an ideological war. Anthropic does not
• 11:02
  need the Pentagon's money to survive or even
• 11:05
  to dominate the market. They clearly have plenty
• 11:07
  of money. Yeah. What they are fighting for is
• 11:09
  legal precedent. They want the courts to affirm
• 11:12
  that a private company can enforce its own safety
• 11:14
  ethics without the government stepping in and
• 11:17
  branding them a national security threat as retaliation.
• 11:19
  And the government, meanwhile, wants to prove
• 11:21
  that when push comes to shove, the state is higher
• 11:24
  on the food chain than the corporation. Exactly.
• 11:27
  It is a fundamental battle over the hierarchy
• 11:29
  of power in the 21st century. Wow. Which brings
• 11:32
  us to our third story and, honestly, the ultimate
• 11:35
  workaround. Yeah, let's look closer at what Metta
• 11:37
  is doing. Because if you're an AI company and
• 11:39
  you can't get enough physical electricity to
• 11:42
  build more data centers like OpenAI in the UK,
• 11:45
  and the government is fighting you over how your
• 11:48
  models are controlled, like Anthropic, what do
• 11:51
  you do? You change how the machine actually thinks.
• 11:54
  Exactly. You find a way to make the software
• 11:56
  fundamentally more efficient, so it doesn't need
• 11:59
  as much electricity or physical infrastructure
• 12:01
  in the first place. And that is exactly what
• 12:03
  Meta has done. This is such a massive leap in
• 12:06
  underlying capability. Meta Superintelligence
• 12:09
  Labs, which is operating under Alexander Wang,
• 12:11
  has just released their newest model, MuseSpark.
• 12:15
  And this is a major comeback moment for Mark
• 12:17
  Zuckerberg's AI ambitions, right? Especially
• 12:19
  after their Llama 4 model kind of stumbled with
• 12:21
  the market back in 2025. Yeah, Llama 4 was a
• 12:24
  bit of a disappointment, but MuseSpark is rolling
• 12:27
  out right now across the US. It's integrating
• 12:29
  directly into WhatsApp, Instagram, Facebook,
• 12:32
  and crucially, Meta's smart glass. Right, and
• 12:35
  it is natively multimodal, meaning it processes
• 12:37
  not just text, but visual data from the physical
• 12:40
  world around you. But the real breakthrough here
• 12:42
  is something Meta is calling Contemplating Mode.
• 12:45
  Yeah, let's talk about this because this is wild.
• 12:48
  It's a total paradigm shift in how AI computes,
• 12:51
  because traditionally when you ask an AI a question,
• 12:54
  it generates a single stream of text. essentially
• 12:57
  predicting the next word in a sequence almost
• 12:59
  instantly. Right, like an autocomplete on steroids.
• 13:02
  Exactly. But contemplating mode changes the architecture
• 13:06
  entirely. It allows the model to orchestrate
• 13:08
  multiple distinct AI agents that reason through
• 13:12
  a problem in parallel. Meta says it can use up
• 13:15
  to 16 agents thinking in concert simultaneously.
• 13:18
  Up to 16 agents. And let's break down how they
• 13:21
  actually trained this, because the mechanism
• 13:23
  they used is fascinating. They used something
• 13:25
  called thinking time penalties. Yes. And to understand
• 13:28
  this, you really need to understand what a token
• 13:30
  is. Think of a token as a fragment of a word,
• 13:34
  or a single piece of a concept. Every time an
• 13:36
  AI generates a token, it consumes computing power,
• 13:40
  which, as we discussed earlier with OpenAI, means
• 13:42
  it consumes actual electricity and physical space
• 13:45
  on a server. So generating a long rambling answer
• 13:48
  isn't just annoying for the user, it is physically
• 13:50
  expensive for the company. Exactly. It costs
• 13:53
  money and power. So during the reinforcement
• 13:56
  learning phase of MuSpark, Meta introduced this
• 13:59
  thinking time penalty. They essentially forced
• 14:02
  the AI to play a high -stakes balancing game.
• 14:05
  Like a trade -off. Right. The model is rewarded
• 14:08
  for getting the mathematically or logically correct
• 14:11
  answer, but it is actively penalized for every
• 14:14
  extra token it uses to get there. So it forces
• 14:17
  the neural network to stop babbling. and start
• 14:20
  building highly efficient compressed pathways
• 14:22
  to the truth. Exactly. Imagine you ask a really
• 14:25
  complex question to a room full of 16 researchers.
• 14:29
  Instead of the loudest person just blurting out
• 14:31
  the first thought that comes to their mind and
• 14:33
  rambling for 10 minutes. Which is what older
• 14:35
  AI models did. Right. Instead of that, the team
• 14:37
  immediately goes into a silent huddle. They debate
• 14:39
  the logic, they verify the math among themselves,
• 14:42
  and then they turn around and hand you a perfectly
• 14:45
  distilled, highly accurate one -sentence answer.
• 14:48
  That is exactly the phase transition meta observed
• 14:51
  during their testing on the AMAD 2025 benchmark.
• 14:54
  The model literally learned to compress its reasoning.
• 14:57
  Wow. And because it's doing this in parallel
• 14:59
  across those 16 agents, the latency, you know,
• 15:03
  the time you spend waiting for the answer remains
• 15:05
  incredibly low. Which means they're getting vastly
• 15:08
  superior intelligence without requiring an exponentially
• 15:11
  larger physical data center. They were solving
• 15:13
  the electricity bottleneck with cognitive efficiency.
• 15:17
  And the real -world applications of this multimodal
• 15:19
  parallel reasoning are moving right into everyday
• 15:22
  life for families and working professionals.
• 15:24
  Yeah, the health applications alone are incredible.
• 15:27
  Metta actually partnered with over 1 ,000 physicians
• 15:30
  to curate health -related training data. Right.
• 15:33
  So because this model is in your smart glasses,
• 15:36
  you can look at a plate of food and the AI can
• 15:38
  estimate the calorie count. Or it can map your
• 15:41
  muscle activity by analyzing your movements during
• 15:43
  a workout. Or visually troubleshoot a broken
• 15:46
  appliance in your kitchen in real time just by
• 15:48
  looking at it. But, you know, the underlying
• 15:50
  logic of that thinking time penalty is what really
• 15:52
  sticks with me. Yeah. Because if Meta is penalizing
• 15:55
  the AI for wasting compute, the model has to
• 15:58
  develop a form of cognitive triage. It has to
• 16:02
  decide, on the fly, is this a simple question
• 16:04
  that I should answer immediately to save electricity,
• 16:07
  or is this a complex problem that requires me
• 16:09
  to spin up 16 agents to debate the logic? It
• 16:13
  is learning to evaluate at own internal resource
• 16:15
  allocation. It is a level of software intuition
• 16:18
  that fundamentally bypasses the brute force physical
• 16:21
  scaling we've seen over the last three years.
• 16:23
  It's honestly incredible. But before we wrap
• 16:25
  up, A quick thanks to Old Glory, an iconic music
• 16:28
  and sports fan merch store. Perfect for the techie
• 16:31
  in your life. Use promo code SNARFL for 15 %
• 16:34
  off at oldglory .com. So bringing all these threads
• 16:38
  together, today's sources really pin a picture
• 16:41
  of a technological future that is incredibly
• 16:43
  messy, deeply constrained, and evolving way faster
• 16:46
  than our physical world can handle. We're watching
• 16:49
  OpenAI hit the hard physical limits of power
• 16:52
  grids and regulatory tape in the UK. proving
• 16:55
  that the digital cloud still very much needs
• 16:57
  dirt, water, and copper wire. And we're watching
• 17:00
  anthropic tangle with the highest levels of the
• 17:03
  US military over who actually gets to dictate
• 17:06
  safety, ethics, and operational control when
• 17:09
  these systems are deployed in the real world.
• 17:11
  And simultaneously, we are watching meta evolve
• 17:14
  AI into multi -agent, contemplating entities
• 17:17
  that are about to live right inside the social
• 17:19
  media apps on your phone and the smart glasses
• 17:21
  on your face. And I think that's why everyone
• 17:23
  listening should care about this. This isn't
• 17:25
  just abstract theory for tech engineers. This
• 17:28
  is about whether the infrastructure of your future
• 17:30
  can even be built. It's about who controls it.
• 17:33
  And it's about how the devices in your pocket
• 17:35
  are actively learning to reason in parallel.
• 17:38
  Absolutely. And I want to leave you with a final
• 17:39
  thought to mull over on your own, connecting
• 17:41
  all three of these massive shifts. OK, let's
• 17:44
  hear it. If models like Meta's Mu Spark are evolving
• 17:46
  to use multiple hidden agents, debating and contemplating
• 17:50
  in parallel to solve complex problems in a fraction
• 17:52
  of a second, and the Pentagon is demanding totally
• 17:56
  unrestricted, unguardrailed access to these kinds
• 17:59
  of models for national security, what happens
• 18:03
  when an unrestricted multi -agent AI is tasked
• 18:06
  with a live military strategy? Wow. That is the
• 18:11
  ultimate question. If 16 AI agents are silently
• 18:14
  debating the logistics of a drone strike in milliseconds,
• 18:17
  will the human commander even be able to follow
• 18:19
  the machine's internal logic before the trigger
• 18:21
  is pulled? It's something we're going to have
• 18:23
  to figure out very soon. Keep questioning the
• 18:25
  systems around you, and thanks for joining us
• 18:27
  as we unpack the reality behind the hype. We're
• 18:30
  here to help at Snarfl, so reach out on our website
• 18:31
  if you have any more questions. That's a wrap
• 18:37
  on today's top stories. Remember, tech doesn't
• 18:40
  have to be complicated, and we're here to help
• 18:42
  connect the dots. If you've got questions or
• 18:45
  want to learn more, visit us at snarful .com.
• 18:48
  Thanks for listening, and we'll catch you next
• 18:50
  time.