Latest / Connecting the Dots / Anthropic's Compute Boom, AI Productivity Puzzles, and Broadcom's Rise

Transcript

Anthropic's Compute Boom, AI Productivity Puzzles, and Broadcom's Rise · Connecting the Dots

• 0:00
  Today's tech news moves fast, but understanding
• 0:03
  it shouldn't be hard. That's why Snarful 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, broadcasting
• 0:17
  directly from the Snarful Solutions Group right
• 0:20
  here in Sacramento, California. I'm Alex. And
• 0:22
  I'm Morgan. And today is Tuesday, April 7, 2026.
• 0:28
  We've got a truly wild deep dive lined up for
• 0:30
  you today. Yeah, we really do. It's it's fascinating
• 0:33
  how much the tech landscape is shifting underneath
• 0:36
  our feet right now. It really is. But before
• 0:38
  we get into the heavy data, let's kick things
• 0:40
  off with our morning star briefing. Just looking
• 0:43
  at the studio window right now, it is honestly
• 0:45
  a beautiful morning here in Sacramento. Oh, absolutely
• 0:48
  stunning. We're looking at mostly sunny skies,
• 0:50
  a really perfect 78 degrees and just this nice
• 0:53
  light breeze of light. two to six miles per hour.
• 0:57
  Super comfortable. Although we do need to point
• 0:58
  out a pretty stark contrast for any of you tuning
• 1:01
  in from outside California today. Right, yeah.
• 1:04
  If you're commuting across the country, it might
• 1:05
  be a much rougher morning. There are currently
• 1:08
  408 active U .S. weather alerts right now. Wow,
• 1:12
  408. Yeah, and 103 of those are actually over
• 1:15
  land. So if you are traveling or just heading
• 1:17
  out the door in those affected regions, please
• 1:20
  do yourself a favor and check weather .gov for
• 1:22
  the latest safety details in your specific area.
• 1:25
  Definitely stay safe out there. Let's pivot over
• 1:27
  to the financial markets, which started the week
• 1:30
  on a decidedly sunny note as well. Looking at
• 1:32
  Monday's close, the Dow Jones jumped up 197 points.
• 1:36
  Nice, closing at 46 ,669 .88. Exactly. And the
• 1:41
  S &P 500 also caught a nice updraft, bumping
• 1:44
  up 24 points to close Monday at 6 ,611 .83. Though
• 1:49
  looking at the digital asset side this Tuesday
• 1:51
  morning, Bitcoin has actually taken a slight
• 1:53
  dip. Oh really? Yeah, it's down 357 points right
• 1:56
  now, just sort of hovering at $68 ,502 .76. Got
• 2:00
  it. Well, let's look closer at our main topic
• 2:02
  today because the internal numbers we are seeing
• 2:04
  paint this really vivid picture of a multi -billion
• 2:07
  dollar arms race. An arms race that is completely
• 2:09
  contrasting with like this massive human behavioral
• 2:12
  experiment happening in the workplace. It's wild.
• 2:15
  Today's episode is brought to you by Strong Coffee,
• 2:18
  the premium instant coffee that tastes like a
• 2:20
  traditional brew, but is packed with 15 grams
• 2:23
  of protein, MCTs, and adaptogens like ashwagandha.
• 2:26
  Get clean, sustained energy without the jitters.
• 2:30
  Use promo code SNARFL for 20 % off at strongcoffeecompany
• 2:34
  .com. So, to really understand the global picture
• 2:38
  of technology right now in 2026, we have to look
• 2:41
  at the physical reality of what Anthropic is
• 2:43
  doing. Right. Anthropic, they're the public benefit
• 2:45
  corporation behind the Claude AI models, right?
• 2:47
  Exactly. And they just finalized this massive
• 2:50
  infrastructure deal with Google and Broadcom
• 2:52
  that fundamentally alters the scale of computing
• 2:55
  as we know it. altering the scale, how? Like,
• 2:58
  what kind of numbers are we talking about here?
• 2:59
  Well, they are securing roughly 3 .5 gigawatts
• 3:02
  of next -generation TPU computing capacity, and
• 3:06
  that's going to start coming online in 2027.
• 3:08
  OK, we definitely need to put the financial momentum
• 3:10
  behind this into perspective first, because the
• 3:12
  revenue growth driving this is what's truly staggering
• 3:15
  to me. Oh, for sure. The financial velocity is
• 3:17
  insane. Right. According to these internal reports
• 3:20
  we pulled, Anthropics run rate revenue has skyrocketed
• 3:24
  to over 30 billion dollars. 30 billion. It's
• 3:27
  just it's hard to even wrap your head around
• 3:28
  that. And to be clear for everyone listening,
• 3:31
  a run rate is just taking their current monthly
• 3:33
  revenue and extrapolating it out over a full
• 3:36
  year. Right. It's a projection. Exactly. But
• 3:38
  at the end of 2025, which, you know, was just
• 3:41
  a few months ago, that run rate was sitting at
• 3:43
  only nine billion dollars. So they have more
• 3:46
  than tripled their financial velocity in less
• 3:49
  than half a year. Yes. And the speed of that
• 3:51
  leap is almost entirely driven by enterprise
• 3:54
  adoption. When they announced their Series G
• 3:57
  fundraising back in February, they had over 500
• 3:59
  business customers who are each spending over
• 4:02
  a million dollars annually. A million bucks a
• 4:04
  year just on AI software. Right, and today. Yeah.
• 4:07
  That number has completely doubled. They now
• 4:10
  have over a thousand enterprise customers dropping
• 4:12
  over a million dollars a year. Which perfectly
• 4:14
  explains this desperate rush for hardware. I
• 4:17
  mean, the demand from corporate America to integrate
• 4:20
  Claude into their daily operations is exploding
• 4:22
  faster than anyone modeled. Definitely. And Broadcom
• 4:25
  is right in the middle of this, right? Oh, absolutely.
• 4:27
  Broadcom CEO, Hawk Tan, noted that they are already
• 4:31
  providing one gigawatt of compute to Anthropic
• 4:33
  right now in 2026. And they're using Google's
• 4:37
  homegrown tensor processing units for that, the
• 4:39
  TPUs. Yeah, exactly. But the demand is surging
• 4:42
  so fast that Anthropic needs to expand from that
• 4:46
  1 gigawatt to 3 .5 gigawatts by next year. And
• 4:50
  from a purely financial perspective, Broadcom
• 4:52
  expects to make $42 billion in AI revenue from
• 4:56
  Anthropic alone by 2027. 42 billion. But let's
• 4:59
  let's actually focus on that 3 .5 gigawatt figure
• 5:02
  for a second, because I feel like gigawatts of
• 5:04
  compute can easily sound like. abstract sci -fi
• 5:07
  jargon. Oh totally I actually need a real world
• 5:09
  translation here. We throw the term gigawatt
• 5:11
  around but what does that actually mean in terms
• 5:13
  of physical grid power? Okay so one gigawatt
• 5:16
  of electricity is roughly the amount of power
• 5:18
  it takes to run an entire medium -sized city.
• 5:20
  Wait really? A whole city? Yeah or like roughly
• 5:24
  the total output of a large nuclear reactor.
• 5:27
  So Anthropic securing 3 .5 gigawatts means they
• 5:31
  are effectively commandeering the electrical
• 5:33
  equivalent of multiple major cities, and they're
• 5:36
  dedicating every single volt of that energy entirely
• 5:39
  to thinking machines. That is just that is unfathomable.
• 5:42
  So we are talking about laying. thousands of
• 5:45
  miles of high voltage cables and building these
• 5:48
  massive industrial cooling systems. Right, literally
• 5:50
  pouring concrete on a scale that rivals the construction
• 5:53
  of the Hoover Dam. And a massive portion of this
• 5:55
  new 3 .5 gigawatt infrastructure is actually
• 5:58
  going to be built right here in the United States,
• 6:00
  which expands on their recent $50 billion commitment
• 6:03
  to American computing. Yeah, the domestic investment
• 6:06
  is huge. But beyond the sheer physical scale,
• 6:09
  what fascinates me is the strategy behind whose
• 6:11
  hardware they are actually buying. Because historically,
• 6:15
  the entire AI boon has basically been a monopoly,
• 6:18
  hasn't it? Oh, completely. For the last several
• 6:19
  years, the entire industry was completely dependent
• 6:22
  on just one company, NVIDIA. Right, the graphics
• 6:25
  processing unit guys. Exactly. OpenAI, Meta,
• 6:28
  Anthropic, they all relied heavily on NVIDIA's
• 6:31
  high -performance GPUs. So if NVIDIA had a bad
• 6:34
  quarter, everyone had a bad quarter. Pretty much.
• 6:37
  If Nvidia faced a manufacturing delay or, say,
• 6:41
  a global supply chain disruption limited the
• 6:43
  availability of those specific chips, the entire
• 6:46
  development of artificial intelligence globally
• 6:48
  would just grind to a halt. Is this multi -platform
• 6:51
  strategy basically like an airline diversifying
• 6:55
  its fleet between Boeing and Airbus? So if one
• 6:59
  manufacturer has a massive delay, the whole airline
• 7:02
  doesn't get grounded? Yes. That is exactly what
• 7:05
  it is. The fundamental goal here is avoiding
• 7:07
  a single point of failure. Which makes a lot
• 7:09
  of sense when you have a thousand enterprise
• 7:11
  customers relying on you. Right. Think about
• 7:13
  it. If an enterprise customer say a major hospital
• 7:16
  network that relies on CLAWD to process patient
• 7:18
  data, or a global logistics firm using it to
• 7:21
  route shipping containers if they need this technology
• 7:23
  for critical second -by -second operations, they
• 7:26
  simply cannot afford for the AI to go down. They
• 7:28
  need guarantees. Exactly. So by expanding away
• 7:32
  from just NVIDIA chips to include Amazon's AWS
• 7:35
  Tranium chips and Google's custom -built TPUs,
• 7:38
  Anthropic is ensuring that the everyday tools
• 7:40
  working professionals rely on won't suddenly
• 7:43
  crash just due to a single supply chain bottleneck.
• 7:46
  It actually makes Claude the only frontier AI
• 7:48
  model available to customers on all three of
• 7:51
  the world's largest cloud platforms, right? Amazon
• 7:53
  Web Services, Google Cloud, and Microsoft Azure.
• 7:57
  Yeah, they are literally matching specific computing
• 8:00
  workloads to the physical chips best suited for
• 8:03
  them. And it's a trend defining the entire sector
• 8:05
  right now. Even OpenAI is doing this, right?
• 8:07
  Yeah, and Thrompet's biggest rival is collaborating
• 8:09
  with Broadcom on custom silicon, too. And they've
• 8:12
  committed to using six gigawatts of AMD's GPUs.
• 8:15
  Wow. You also have massive independent projects
• 8:18
  like TerraFab, which is this vertically integrated
• 8:21
  chip manufacturing facility backed by Elon Musk's
• 8:24
  ventures trying to create more alternatives.
• 8:26
  So Nvidia is still the dominant force, obviously,
• 8:28
  but the market is actively funding. credible
• 8:31
  alternatives to build out supply chain resilience.
• 8:34
  Well, moving on to our next story, let's look
• 8:35
  at what happens when all of that unfathomable
• 8:38
  city -powering compute actually reaches the desk
• 8:41
  of the everyday corporate worker. Oh man, this
• 8:44
  is where it gets crazy. Because while the structural
• 8:46
  engineers at Broadcom and Google are sweating
• 8:49
  over global supply chains and nuclear -level
• 8:51
  energy grids, the employees over at Metta have
• 8:54
  turned AI usage into a literal video game. The
• 8:58
  contrast is genuinely jarring, but it raises
• 9:01
  a really vital question about how corporate America
• 9:04
  actually measures productivity in the AI era.
• 9:07
  Okay, break this down for us because this leaked
• 9:09
  data is wild. So internal sources leaked details
• 9:13
  about an intranet leaderboard that Meta has set
• 9:15
  up for its 85 ,000 employees. And this leaderboard
• 9:18
  is internally dubbed Claudionomics. Claudionomics.
• 9:21
  I mean, you cannot make this stuff up. I know,
• 9:23
  right? And the explicit purpose of this leaderboard
• 9:25
  is to track AI token consumption across the entire
• 9:28
  workforce. Wait, so employees are actively competing
• 9:30
  to see who can consume the most AI computing
• 9:32
  power. Yes. In a single 30 -day period, meta
• 9:35
  -employees burn through 60 trillion tokens. Hold
• 9:39
  on. 60 trillion. For anyone outside the engineering
• 9:43
  department, what physically is a token in this
• 9:45
  context? Because 60 trillion of anything sounds
• 9:48
  incredibly expensive. It is. A token is the fundamental
• 9:52
  unit of compute for these models. Think of it
• 9:54
  as like a piece of a word that the AI processes.
• 9:58
  Okay. So every single time an employee asks an
• 10:01
  AI a question or has a read a PDF document or
• 10:04
  asks it to generate a block of code, the system
• 10:07
  burns tokens to process that request. And crucially,
• 10:11
  every single token costs real money. Real money
• 10:13
  and physical electricity from those gigawatt
• 10:15
  data centers we just talked about. But the leaderboard
• 10:17
  has heavily gamified this expense. It uses these
• 10:20
  role -playing game titles to get employees hooked
• 10:22
  on utilizing AI tools. Oh, the titles are the
• 10:25
  best part. Depending on how much raw compute
• 10:27
  you burn through, you can earn titles like Cash
• 10:29
  Wizard. Cash Wizard. Or Model Connoisseur. Yeah.
• 10:31
  But the ultimate top tier rank is Token Legend.
• 10:34
  Token Legend. Yeah. And the highest ranking user
• 10:36
  on this leaderboard, the ultimate token legend,
• 10:39
  averaged 281 billion tokens in a single 30 -day
• 10:43
  window. One single user. This behavior has spawned
• 10:47
  this bizarre new trend across Silicon Valley
• 10:49
  known as token maxing. Yeah, token maxing is
• 10:52
  out of control. Explain the mechanics of how
• 10:55
  an employee actually achieves that. because no
• 10:58
  single human is manually typing out enough prompts
• 11:00
  to generate 281 billion tokens in a month. Oh,
• 11:04
  they don't type it out. They automate the waste.
• 11:06
  Automate the waste. Token maxing is the practice
• 11:09
  of leaving AI agents running autonomously for
• 11:12
  hours or even days on end. They have them perform
• 11:15
  highly redundant, complex, or honestly just pointless
• 11:18
  tasks solely to pad the employees' numbers on
• 11:22
  the leaderboard. That is absurd. And it gets
• 11:24
  worse when you factor in what they call reasoning
• 11:26
  tokens. What are reasoning tokens? So newer AI
• 11:30
  models are designed to essentially talk to themselves
• 11:32
  to puzzle through a problem before they actually
• 11:34
  give the user an answer. Oh, right. So they show
• 11:36
  their work, basically. Exactly. So an employee
• 11:39
  can set an AI to solve a complex puzzle, and
• 11:41
  the AI burns millions of tokens just debating
• 11:44
  with itself. And the employee's leaderboard rank
• 11:46
  shoots straight up. Regardless of whether the
• 11:48
  final answer actually helped the company do anything
• 11:51
  valuable. Precisely. The crazy part to me is
• 11:54
  that this isn't just a grassroots phenomenon
• 11:56
  of board engineers playing games. It is actively
• 11:59
  being driven from the top down. Oh, yeah. Leadership
• 12:02
  loves this. Right. NVIDIA CEO Jensen Huang recently
• 12:06
  went on record stating that he would be deeply
• 12:09
  alarmed if an engineer making $500 ,000 a year
• 12:13
  wasn't consuming at least $250 ,000 worth of
• 12:16
  compute tokens. And you also have Metta's CTO,
• 12:19
  Andrew Bosworth, supposedly bragging that one
• 12:21
  of his top engineers spends the equivalent of
• 12:23
  his entire salary on tokens, claiming it increased
• 12:26
  his output by 10x. Okay, I have to call a timeout
• 12:29
  on this corporate logic. This is like judging
• 12:30
  a delivery driver by how much gas they waste
• 12:32
  instead of how many packages they deliver. That
• 12:34
  is a great way to put it. Or tracking a writer's
• 12:36
  productivity by how many ink cartridges they
• 12:38
  empty into the trash. It just proves the engine
• 12:40
  is on, not that you're going anywhere. Exactly.
• 12:43
  The fundamental flaw in this metric is mistaking
• 12:46
  activity for value. Every one of those 60 trillion
• 12:50
  tokens represents a hard financial cost to Meta.
• 12:53
  Has anybody actually proved that this works?
• 12:56
  No. Nobody has produced hard data proving that
• 13:00
  burning all these tokens translates directly
• 13:02
  to real business revenue. So they're just guessing.
• 13:05
  For AI companies, finding the connection between
• 13:08
  raw usage and actual profit generation is the
• 13:11
  single most critical challenge they face right
• 13:13
  now. I mean, they desperately have to justify
• 13:16
  the billions of dollars and the gigawatts of
• 13:18
  power they are pouring into this infrastructure.
• 13:21
  It feels like a massive bubble of self -validation.
• 13:24
  And we've seen this tactic before. Google used
• 13:26
  to routinely report raw token consumption in
• 13:29
  its cloud offerings during quarterly earnings
• 13:31
  calls. Just to show Wall Street investors that
• 13:33
  people were theoretically adopting the technology.
• 13:36
  Right. Showing raw usage numbers looks stratospheric
• 13:39
  on a slide deck, but the actual value delivered
• 13:42
  to the human user might be entirely marginal.
• 13:44
  Exactly. If a developer uses a million tokens
• 13:47
  to generate a piece of code that doesn't actually
• 13:49
  work, the leaderboard still calls them a cash
• 13:51
  wizard, but the company just lost money. Totally
• 13:54
  backwards. Relying on consumption as a proxy
• 13:56
  for productivity is a strategy that simply won't
• 13:59
  survive long -term investor scrutiny. Before
• 14:02
  we wrap up, a quick thanks to Old Glory, an iconic
• 14:05
  music and sports fan merch store, perfect for
• 14:08
  the techie in your life. Use promo code SNARFL
• 14:11
  for 15 % off at OldGlory .com. So, to bring these
• 14:16
  two massive shifts into focus for you, we are
• 14:19
  watching a fascinating dichotomy unfold in real
• 14:23
  time right now. It really is a tale of two extremes.
• 14:25
  Right. On the hardware and infrastructure side,
• 14:28
  companies like Anthropic, Broadcom, and Google
• 14:30
  are engaging in an incredibly serious multi -billion
• 14:34
  dollar arms race. Doing the heavy lifting. They
• 14:36
  are successfully building the biggest, most resilient,
• 14:39
  most power -hungry AI supercomputers in human
• 14:42
  history. They are literally redrawing the map
• 14:45
  of American energy consumption to ensure global
• 14:47
  supply chains don't collapse. Yet when that unimaginable
• 14:50
  city -level power trickles down to the actual
• 14:53
  office workers, corporate management hasn't quite
• 14:55
  figured out what to do with it yet. Not at all.
• 14:57
  We are currently just turning it into a ridiculous
• 14:59
  video game of who can burn the most digital fuel
• 15:02
  at companies like Metta. It is a profound mismatch
• 15:06
  between the raw capability of the tool and the
• 15:09
  immaturity of the metrics we use to measure its
• 15:12
  actual impact. We have successfully built the
• 15:14
  fastest, most expensive race car in the world,
• 15:17
  but we are still arguing over how to measure
• 15:20
  who is actually winning the race. Which leaves
• 15:22
  us with a pretty mind -bending question for you
• 15:24
  to ponder on your commute or while you're grabbing
• 15:26
  your coffee today. Oh, I like this one. Think
• 15:28
  about the mechanics of token maxing for a second.
• 15:31
  If corporate America firmly decides that AI token
• 15:35
  usage, basically how much compute you can waste
• 15:38
  is the ultimate metric for how valuable you are
• 15:40
  as an employee. What happens when these AI models
• 15:44
  inevitably become drastically more efficient?
• 15:46
  Because as the software naturally improves and
• 15:49
  optimizes, it will require significantly fewer
• 15:51
  tokens to accomplish the exact same task. Exactly.
• 15:55
  Will workers suddenly be penalized or look less
• 15:57
  productive during their performance reviews?
• 16:00
  Simply because the AI learned how to do the exact
• 16:02
  same job using half the computing power. It creates
• 16:05
  the ultimate efficiency paradox. If your professional
• 16:09
  worth is tied to your digital waste, true innovation
• 16:12
  might actually cost you your title as a token
• 16:15
  legend. We're here to help at Snarful, so reach
• 16:17
  out on our website if you have any more questions.