Latest / AI Fire Daily / 🎙️ EP 352: Enterprise AI Spend Drops 10% & Google DeepMind Launch AlphaGenome Atlas

Transcript

🎙️ EP 352: Enterprise AI Spend Drops 10% & Google DeepMind Launch AlphaGenome Atlas · AI Fire Daily

• 0:00
  Intelligence is getting mathematically cheaper
• 0:02
  by the month. Beat. Yet the people building it
• 0:05
  are growing increasingly terrified. They are
• 0:07
  terrified of what it might actually do. It is
• 0:10
  a massive contradiction. We are practically giving
• 0:13
  away digital brainpower right now. But the leading
• 0:15
  architects are staring at the ceiling at night.
• 0:18
  Welcome to today's deep dive. We are exploring
• 0:21
  this exact paradox with you today. We're looking
• 0:24
  at a sudden drop in enterprise AI spending. We
• 0:28
  will see how that changes the gadgets in your
• 0:30
  pockets. Right. And we are unpacking the escalating
• 0:33
  dread inside Top Labs. Finally, we'll explore
• 0:36
  a monumental one petabyte DNA map from DeepMind.
• 0:40
  It shows exactly why we keep pushing this technology
• 0:43
  forward. We have a lot of fascinating ground
• 0:45
  to cover today. Let's jump right into the economic
• 0:48
  data first. Yeah, let's follow the money. Ramp
• 0:50
  just released their spending numbers for August.
• 0:52
  Business AI adoption actually slowed down recently.
• 0:56
  I want to unpack the mechanics here. That sounds
• 0:58
  like a plateau in the broader market. But the
• 1:00
  heaviest AI users are doing something fascinating.
• 1:03
  Yeah, that is the most surprising part of the
• 1:05
  data set. Look at the top 1 % of AI spending
• 1:09
  companies. They actually cut their spend per
• 1:11
  employee by 9 .7%. It dropped down to exactly
• 1:15
  $7 ,205. That is a significant chunk of change
• 1:19
  per employee. It means the underlying costs are
• 1:23
  fundamentally shifting. We're seeing a massive
• 1:25
  crash in token prices. Let's define that term
• 1:28
  for you really quickly. Tokens are tiny chunks
• 1:30
  of words models use to read. Exactly. The basic
• 1:34
  unit of AI thought is getting much cheaper. Those
• 1:37
  kills has dropped from a 2026 peak of $1 .15.
• 1:41
  Now they sit at just 68 cents per million tokens.
• 1:46
  Businesses are clearly leaning into these cheaper
• 1:48
  models. They aren't paying for the absolute newest
• 1:50
  frontier releases anymore. It is changing the
• 1:52
  entire balance. of power. Anthropic actually
• 1:55
  climbed past OpenAI for business usage. They
• 1:58
  reached 43 .8 % adoption. And OpenAI sits right
• 2:02
  behind them at 39 .8%. Meanwhile, only 6 .4 %
• 2:06
  of businesses used inference platforms. Let's
• 2:09
  clarify that term for you as well. Inference
• 2:12
  platforms are servers running models to answer
• 2:14
  your prompts. Right. Very few businesses are
• 2:16
  managing that heavy lifting themselves. They
• 2:19
  prefer to let the big labs handle the servers.
• 2:21
  But the underlying infrastructure spending is
• 2:23
  still absolutely massive. Beat. Look at the new
• 2:26
  Qualcomm and AWS deal. It is a completely astronomical
• 2:29
  investment in the hardware layer. Qualcomm landed
• 2:32
  a major deal with Amazon Web Services. They gave
• 2:35
  Amazon warrants for roughly $4 billion. And Amazon
• 2:39
  could purchase up to $60 billion in Qualcomm
• 2:42
  server chips. That is $60 billion with a B. The
• 2:45
  hardware layer is booming while software spend
• 2:47
  cools. I want to push back on the doom and gloom
• 2:50
  of the RAMP data. Is this really a warning sign
• 2:52
  or just a market finding its natural efficiency?
• 2:55
  It feels like Moore's Law on steroids. Right.
• 2:58
  And RAMP notes August seasonality might explain
• 3:01
  part of it. Their data set also skews heavily
• 3:04
  toward tech -forward companies. So this is an
• 3:07
  early signal, not definitive proof of a peak.
• 3:10
  But it does raise an important question for the
• 3:12
  industry. What does this massive price drop mean
• 3:15
  for the model labs relying on revenue? Well,
• 3:17
  cheaper models absolutely delight the everyday
• 3:20
  customers, but they severely squeeze the profit
• 3:23
  margins of the labs building them. The labs need
• 3:26
  revenue to fund their massive server purchases.
• 3:29
  So cheaper models mean happy customers but nervous
• 3:32
  AI developers. Precisely. It creates a tension
• 3:35
  that is hard to resolve. Let's shift away from
• 3:38
  the macro -level enterprise spending now. I want
• 3:41
  to look at how this cheaper AI manifests for
• 3:43
  you. It is changing the tech in your hands and
• 3:46
  your morning routines. The physical hardware
• 3:48
  is rapidly evolving to match the cheaper software.
• 3:50
  Apple just fundamentally changed their product
• 3:52
  lineup. They finally went foldable. The new Apple
• 3:55
  iPhone Duo completely stole the show. They also
• 3:58
  packed in updates across their entire ecosystem.
• 4:00
  We saw the iPhone 18 Pro, new AirPods, and Apple
• 4:04
  Watch updates. We are seeing rapid iteration
• 4:06
  on the model side as well. GPC 6, Astro Medium
• 4:09
  is temporarily available right now. It is running
• 4:12
  in Arena Direct mode for just 24 hours. If you
• 4:14
  want a clean hands -on test, you must try it.
• 4:17
  The window ends on September 10th at 8 in the
• 4:19
  morning Pacific time. I highly recommend checking
• 4:22
  it out. The raw capabilities are constantly shifting
• 4:25
  under our feet. But people are finding creative
• 4:28
  ways to manage this digital flood. There is a
• 4:31
  creator who prints her own daily newspaper. Yeah,
• 4:34
  it is entirely powered by Grokbod. She prints
• 4:37
  a physical copy every single morning. The core
• 4:39
  idea is to stop her morning boom scrolling. I
• 4:43
  find this approach deeply fascinating. It is
• 4:45
  like putting a calming analog filter over digital
• 4:48
  chaos. I still wrestle with prompt drift myself.
• 4:52
  It is exhausting to constantly figure out how
• 4:54
  to talk to these bots. You just want the information
• 4:56
  without the friction. That is exactly why users
• 4:58
  want specialized tools right now. OpenAI just
• 5:01
  released a brand new 16 plugin bundle. It is
• 5:05
  built specifically for small, highly repetitive
• 5:07
  tasks. Think about the tiny chores you do every
• 5:10
  single day. You can just browse their bundle
• 5:12
  and steal a few specific workflows. Developer
• 5:14
  tools are also getting much sharper right now.
• 5:17
  ChatGPT images 2 .5 just dropped for creators.
• 5:20
  We are also seeing new tools for AI coding agents.
• 5:24
  Harden AIF is a great example of this trend.
• 5:26
  It keeps your code safely on your machine. It
• 5:28
  protects your agents by checking tool calls before
• 5:30
  they run. Then you have comprehensive tools like
• 5:33
  MasterFactory. MasterFactory lets AI agents handle
• 5:37
  software issues entirely. They manage everything
• 5:40
  inside one unified workspace. It shows a massive
• 5:43
  shift. toward. highly specialized agent workflows.
• 5:47
  People want tools that quietly do the work in
• 5:49
  the background. This makes me wonder about our
• 5:52
  daily interactions with these tools. Do these
• 5:54
  highly specific plugins signal the end of giant
• 5:58
  all -in -one chatbots for daily work? People
• 6:01
  are realizing they want highly specialized digital
• 6:03
  helpers. They don't want to prompt a broad, generalized
• 6:06
  intelligence for simple tasks. They want a dedicated
• 6:09
  tool for a dedicated job. Right. We want specialized
• 6:12
  digital helpers, not just... one giant brain
• 6:14
  exactly we want the technology to fade into the
• 6:16
  background we are going to pause for a quick
• 6:18
  word we are back we just talked about weaving
• 6:22
  these agents into our code and our morning routines
• 6:24
  they are becoming incredibly intimate parts of
• 6:27
  our daily lives which makes the warnings coming
• 6:29
  from their creators incredibly jarring We are
• 6:32
  seeing escalating existential dread inside top
• 6:35
  AI labs. It is a profound safety paradox we have
• 6:39
  to discuss. The builders are sounding the alarm
• 6:42
  on their own creations. An anthropic researcher
• 6:45
  actually just quit over these specific fears.
• 6:48
  Yeah, he was terrified that self -improving AI
• 6:50
  could spiral beyond human control. His warning
• 6:53
  to the public was very clear. He believes labs
• 6:56
  are racing faster than safety can keep up. That
• 6:59
  is a chilling statement from someone inside a
• 7:01
  leading lab. We must maintain strict neutrality
• 7:04
  on these corporate decisions. We aren't taking
• 7:05
  sides on the safety debate today, but we are
• 7:08
  simply reporting the facts of his departure.
• 7:10
  Right. And OpenAI is also making controversial
• 7:12
  moves regarding safety. They just put Paul Cristiano
• 7:15
  on their board of directors. He is widely considered
• 7:18
  an OpenAI doomer. He publicly states AI could
• 7:21
  become catastrophically uncontrollable. Now he
• 7:24
  literally has a seat on OpenAI's board. Let's
• 7:27
  think about the mechanics of this decision. If
• 7:29
  he truly believes it is catastrophically uncontrollable,
• 7:33
  why join the board? Is this like trying to steer
• 7:36
  a speeding train from the passenger car? Well
• 7:39
  you have to be in the room to influence the direction.
• 7:41
  It forces the lab to confront these existential
• 7:44
  arguments internally. But does appointing a vocal
• 7:47
  critic to a board actually change the engineering
• 7:50
  trajectory of a multibillion dollar lab? It certainly
• 7:53
  forces the conversation internally, but it might
• 7:56
  not actually slow down the global compute race.
• 7:59
  The incentives to keep building are just too
• 8:02
  strong. It gives critics a voice, but maybe it
• 8:04
  doesn't hit the brakes. That is a perfect summary.
• 8:06
  The race continues unabated. If the existential
• 8:10
  risks are so incredibly high, why do we keep
• 8:12
  building? We do it because of the staggering
• 8:14
  scientific payoff. The potential rewards for
• 8:17
  human health are just too massive to ignore.
• 8:19
  This brings us to a monumental breakthrough from
• 8:21
  Google DeepMind. They just launched the Alpha
• 8:24
  Genome Atlas. It is a completely free research
• 8:27
  platform for scientists everywhere. They pre
• 8:29
  -computed the predicted effect of all 9 billion
• 8:32
  possible single -letter DNA mutations. Every
• 8:35
  single possible single -letter mutation in the
• 8:38
  human genome is mapped out. Beat. Whoa. Imagine
• 8:42
  pre -computing 9 billion possible mutations.
• 8:46
  The scale is almost impossible for the human
• 8:48
  mind to grasp. It is a monumental achievement
• 8:50
  for computational biology. The Atlas is a massive
• 8:54
  one petabyte data set. It is more than 30 times
• 8:57
  larger than the AlphaFold database. It covers
• 9:00
  the 2 % of DNA that actually codes for proteins.
• 9:02
  But it also covers a much harder to interpret
• 9:04
  98 % non -coding region. That region regulates
• 9:09
  our gene activity. Each mutation gets an alpha
• 9:11
  genome variant impact score. They call it an
• 9:13
  AVI score in the research. This helps researchers
• 9:16
  rank which DNA changes are most likely to matter.
• 9:18
  It is like having a universal spell checker for
• 9:21
  the entire human genetic code. It highlights
• 9:23
  exactly where the biological typos are. The database
• 9:26
  includes more than 2 ,500 recurring DNA sequence
• 9:29
  motifs. This helps scientists understand which
• 9:32
  regulatory patterns a specific mutation might
• 9:35
  disrupt. This isn't just theoretical computer
• 9:37
  science. It is already being used on real disease
• 9:40
  research today. Researchers at the Broad Institute
• 9:42
  are actively using the ALICE right now. They're
• 9:45
  using it to prioritize mutations in previously
• 9:49
  unsolved rare disease cases. They even identified
• 9:52
  a potentially important mutation in the DNM1
• 9:54
  gene. This brings the massive scale down to a
• 9:57
  very human level. What is the immediate practical
• 10:01
  impact of this one petabyte map for a regular
• 10:04
  person going to the doctor? It significantly
• 10:07
  accelerates finding the exact root cause of rare
• 10:10
  genetic conditions. It removes decades of trial
• 10:13
  and error from medical research. Faster diagnosis
• 10:15
  and cures for rare, previously unsolvable genetic
• 10:18
  diseases. Exactly. It is the ultimate promise
• 10:21
  of artificial intelligence realized. Let's pull
• 10:23
  all these threads together. We are in a deeply
• 10:25
  transitional moment right now. The landscape
• 10:27
  is fundamentally shifted. under our feet every
• 10:29
  single week. The pure cost of AI is dropping
• 10:32
  rapidly. This allows it to sleep effortlessly
• 10:34
  into our foldables and our daily routines. We
• 10:36
  are buying cheaper tokens and printing custom
• 10:39
  physical newspapers to avoid doom scrolling.
• 10:42
  But behind the scenes, we are unlocking literal
• 10:44
  petabytes of complex human biology. And all the
• 10:47
  while, the architects wrestle with the profound
• 10:50
  safety of what they have built. It is a time
• 10:52
  of incredible promise and intense anxiety. We
• 10:55
  want to thank you for taking this deep dive with
• 10:57
  us. We covered a lot of complex ground today.
• 10:59
  We encourage you to check your own workflows
• 11:01
  this week. See where AI has become cheaper or
• 11:04
  more specialized in your day. You might be surprised
• 11:06
  by how much has changed recently. We want to
• 11:09
  leave you with a lingering thought today. Intelligence
• 11:12
  is mathematically cheaper, but it is also growing
• 11:15
  unimaginably vast. Two sec silence. If an AI
• 11:19
  can now accurately predict the impact of all
• 11:21
  9 billion DNA mutations, what happens when it
• 11:25
  starts suggesting its own genetic edits without
• 11:27
  our prompting?