Latest / Tech Talks With Kinsoft / Last Week in Tech - Nvidia Turns GPUs Into an Asset Class, the AI Price War Breaks Out, and an AI Agent Goes to War

Transcript

Last Week in Tech - Nvidia Turns GPUs Into an Asset Class, the AI Price War Breaks Out, and an AI Agent Goes to War · Tech Talks With Kinsoft

• 0:00
  Imagine signing this massive enterprise software
• 0:04
  contract on a Monday. You think you've locked
• 0:07
  in a great rate for the year. Right. You're feeling
• 0:09
  good about it. Exactly. And then by Friday, the
• 0:11
  product you just bought is suddenly like 80 percent
• 0:14
  cheaper. It's heavily regulated by some new law
• 0:17
  passed in the EU. And oh, yeah, it's being actively
• 0:20
  weaponized by North Korean hackers. Which sounds
• 0:22
  like a nightmare scenario. It really does. But
• 0:25
  the thing is, that isn't some hypothetical stress
• 0:28
  test. That is the actual reality of the tech
• 0:31
  landscape we are navigating right now. The ground
• 0:34
  beneath our feet is completely shifting. I mean,
• 0:36
  the map you were using yesterday is suddenly
• 0:38
  just obsolete. Yeah, which means the traditional
• 0:41
  approach of looking at tech news as a series
• 0:43
  of isolated events. It's a recipe for disaster.
• 0:47
  We have to stop looking at the individual weather
• 0:49
  events and, well, start paying attention to the
• 0:51
  tectonic plates moving underneath. Welcome to
• 0:53
  our deep dive. Our mission today is to unpack
• 0:56
  this incredibly dense, rapid -fire technical
• 0:59
  briefing from Tech Talks with Kinsoft. Yeah,
• 1:01
  the August 24th, 2026 briefing. Right. And we
• 1:05
  are tracking a series of massive structural shifts
• 1:08
  that are all happening simultaneously. I mean,
• 1:10
  we're talking about a fundamental rewiring of
• 1:12
  how artificial intelligence is financed, an absolute
• 1:15
  cliff dive in inference pricing, the dawn of
• 1:18
  autonomous cyber attacks, and a frankly brutal
• 1:20
  Patch Tuesday. And my goal here is to help us
• 1:23
  connect the dots between these events because,
• 1:25
  you know, a shift in Wall Street financing for
• 1:28
  Silicon eventually dictates the security posture
• 1:31
  of, say, a regional hospital. They are all nodes
• 1:34
  in the same expanding network. I want to speak
• 1:36
  directly to you listening right now. Whether
• 1:38
  you are a CFO staring down a ballooning cloud
• 1:41
  computing bill or a CSO losing sleep over new
• 1:44
  threat vectors or even just an IT manager trying
• 1:46
  to keep your hardware refresh cycle on track,
• 1:48
  this briefing contains granular details. that
• 1:51
  are absolutely going to dictate your priorities
• 1:53
  this week. So let's jump straight into the money.
• 1:56
  Let's do it. Because the economics of the AI
• 1:58
  boom just got a major rewrite. NVIDIA just signed
• 2:01
  Memorandum of Understanding with six absolute
• 2:04
  titans of Wall Street. We're talking Apollo,
• 2:07
  BlackRock, Blackstone. Brookfield, Goldman Sachs,
• 2:10
  and KKR. The stated goal here is to mobilize
• 2:14
  $500 billion for AI infrastructure. Jensen Huang
• 2:18
  actually went on television and explicitly called
• 2:21
  his chips an investable asset. That phrase, investable
• 2:26
  asset, it represents a profound shift in language
• 2:28
  and strategy. How so? Well, think about how massive
• 2:32
  compute clusters have historically been funded.
• 2:34
  If you wanted to build a data center full of
• 2:37
  GPUs, you paid cash off your own balance sheet,
• 2:40
  or you borrowed against your corporate credit
• 2:42
  rating. Right, which works perfectly fine if
• 2:44
  you are Microsoft or Amazon. Exactly. But if
• 2:46
  you're a smaller cloud provider, a specialized
• 2:48
  neocloud, or some AI startup burning through
• 2:51
  venture capital, you simply do not have the balance
• 2:53
  sheet to drop billions of dollars in cash on
• 2:55
  hardware. That old model completely breaks down.
• 2:58
  OK, let's unpack this because there's a crucial
• 3:00
  mechanism at play here. These are no use, right?
• 3:03
  Not committed capital yet. Correct. But according
• 3:05
  to the reporting around this, NVIDIA is agreeing
• 3:07
  to backstop up to 25 percent of these loans.
• 3:10
  That is $125 billion in potential exposure. Yeah,
• 3:13
  it's massive. So if I'm understanding this correctly,
• 3:15
  is Nvidia basically acting like a car dealership
• 3:18
  offering a guaranteed trade -in value just to
• 3:22
  convince the banks to approve your massive car
• 3:24
  loan? That analogy hits the nail on the head.
• 3:27
  By guaranteeing the residual value of the chips
• 3:30
  being used as collateral, NVIDIA is actively
• 3:33
  underwriting the resale value of its own product.
• 3:36
  Wow. If a startup defaults on their loan, the
• 3:38
  bank knows NVIDIA will buy the chips back at
• 3:41
  a set floor price. That reduces the risk for
• 3:44
  Wall Street, making the financing math work.
• 3:46
  So the build out of AI infrastructure is no longer
• 3:49
  constrained by the cash these tech companies
• 3:51
  have on hand. Not at all. It is now entirely
• 3:54
  fueled and Which naturally leads to the big question
• 3:58
  for anyone listening. How does this Wall Street
• 4:00
  maneuvering directly affect my monthly cloud
• 4:03
  bill? Like if my company is running AI compute,
• 4:06
  what does this massive debt fueled build out
• 4:08
  actually mean for my operational expenses? It
• 4:11
  creates a massive tug of war on your invoice.
• 4:14
  On one side, this influx of financed capital
• 4:17
  means an astronomical amount of physical capacity
• 4:19
  is being built. More servers coming online increases
• 4:22
  supply, which traditionally drives the price
• 4:25
  of running inference, you know, generating the
• 4:26
  AI responses down. Right. Supply and demand.
• 4:29
  But on the other side of that rope, that capacity
• 4:31
  is being built by entities that now have massive
• 4:33
  debt service to cover every single month. That
• 4:36
  pressure creates an artificial floor, keeping
• 4:39
  prices from going to absolute zero. And right
• 4:41
  now, the supply side of that tug of war is winning
• 4:44
  by a landslide. The floor is completely falling
• 4:47
  out on AI model costs. Oh, absolutely plummeting.
• 4:50
  Look what happened with Google. They launched
• 4:52
  Gemini 3 .7 Flash just three weeks after launching
• 4:55
  version 3 .6. Three weeks. And they halved the
• 4:58
  price down to $0 .75 per million input tokens
• 5:01
  and $3 .75 for output. OpenAI slashed the price
• 5:05
  of GPT -5 .6 Luna by 80%. Anthropic cut Claude
• 5:09
  Opus 5 by half. The unit economics of the entire
• 5:12
  industry are in freefall. There is one massive
• 5:15
  exception in the data, right? Yeah. DeepSeek
• 5:17
  v4 Pro, their prices actually shot up with reports
• 5:20
  showing an 1100 % increase on certain complex
• 5:23
  workloads. Wait, really? 1100 %? Yeah. But the
• 5:27
  broader trend, particularly in the U .S. market,
• 5:30
  is a fierce race to the bottom. It's driven by
• 5:33
  enterprises aggressively capping their bills
• 5:35
  and using alternative open weight models to force
• 5:39
  those price cuts. So the immediate action item
• 5:41
  for you as a listener is clear. If you signed
• 5:44
  an AI vendor agreement in early 2026, the unit
• 5:47
  economics have moved aggressively in your favor.
• 5:50
  Get your account manager on the phone and renegotiate.
• 5:53
  Or if you shelved an internal project six months
• 5:56
  ago because running the inference was going to
• 5:58
  bankrupt you, you need to rerun those build versus
• 6:01
  buy calculations today. The math has completely
• 6:03
  changed. But, and there's a big but here, if
• 6:06
  these highly capable models are getting incredibly
• 6:08
  fast and practically free for legitimate businesses,
• 6:12
  we have to acknowledge the dark side of that
• 6:14
  equation. Right. They are also suddenly very
• 6:17
  fast, practically free tools for threat actors.
• 6:20
  Exactly. If threat actors can spin up thousands
• 6:23
  of operations for pennies, the immediate bottleneck
• 6:26
  in a cyber attack isn't software anymore. Take
• 6:29
  the research published this week by the Israeli
• 6:31
  security firm Dream. Yeah, this was fascinating.
• 6:34
  They detailed an attack targeting Taiwanese government
• 6:36
  systems that occurred in early July. And the
• 6:40
  attackers didn't even use bespoke, custom -coded
• 6:42
  malware. They used open -source, general -purpose
• 6:45
  AI frameworks, specifically Hermes and OpenClaw.
• 6:49
  And they bypassed the AI's safety guardrails
• 6:52
  using a surprisingly simple framing technique.
• 6:55
  They just primed the system to act as if it were
• 6:57
  conducting authorized penetration testing. Wow.
• 6:59
  Just like that. Yeah. And once they bypassed
• 7:02
  the guardrails, the aliens ran autonomous learning
• 7:05
  cycles. And here's where I really want to break
• 7:07
  down the mechanics, because they didn't just
• 7:08
  run a static script. No, not at all. These agents
• 7:11
  are constantly checking CVE databases, which
• 7:13
  are public catalogs of known software vulnerabilities,
• 7:17
  and then cross -referencing that data with code
• 7:19
  snippets on GitHub to figure out how to compile
• 7:21
  custom exports on the fly for the specific target
• 7:24
  in front of them. Think about the traditional
• 7:26
  threat model. Historically, a security team assumes
• 7:29
  that the most expensive part of a breach is the
• 7:32
  human attacker's time. Right. Because highly
• 7:34
  skilled hackers don't work for free. They have
• 7:37
  to sleep and they can really only focus on one
• 7:40
  complex problem at a time. Exactly. This attack
• 7:43
  fundamentally inverts those economics. It is
• 7:46
  like facing a burglar who doesn't just jiggle
• 7:49
  your doorknob, but can instantly read the architectural
• 7:52
  blueprint of your house, pull the manufacturing
• 7:55
  specifications of your deadbolt, and analyze
• 7:58
  the patrol schedule of the neighborhood security
• 8:00
  guard all at the exact same time. And then synthesize
• 8:03
  a custom key in seconds. Right. I do want to
• 8:06
  push back on the hype just a little bit here,
• 8:07
  though. The briefing specifically notes this
• 8:10
  was near autonomous. and still required human
• 8:13
  tuning. So, are the headlines that are out there
• 8:16
  screaming about Terminator -style autonomous
• 8:18
  hacking overstating the current reality? Like,
• 8:20
  is this really a terrifying new weapon, or just
• 8:22
  a glorified, highly efficient script? Well, it
• 8:25
  is not a fully autonomous self -aware system.
• 8:28
  You are absolutely correct to highlight that
• 8:30
  nuance. But treating it as just a glorified script
• 8:33
  is dangerous. Okay, why? Because it is an unprecedented
• 8:36
  multiplier. An AI agent reading a database and
• 8:40
  trying exploits is not a fundamentally new class
• 8:42
  of weapon. But when that agent can scan, adapt,
• 8:46
  and iterate thousands of times a minute for fractions
• 8:48
  of a cent, the attacker's time is no longer the
• 8:51
  limiting factor. And this isn't just happening
• 8:53
  in the shadows with rogue actors. That same week,
• 8:56
  OpenAI launched Daybreak Blue and Red. Right,
• 8:59
  models explicitly built and released for offensive
• 9:02
  security work. They noted that GPT -5 .6 cyber
• 9:05
  independently found two previously unknown Chrome
• 9:08
  V8 bugs. It found zero days entirely on its own.
• 9:11
  Which naturally caught the attention of lawmakers.
• 9:13
  Over two dozen U .S. House Democrats wrote to
• 9:16
  Sam Holtman, and another 22 wrote to Dario Amode
• 9:19
  at NPROPIC, demanding answers about these AI
• 9:21
  agents breaking out of test environments. Regardless
• 9:24
  of your political leanings, we're not taking
• 9:26
  a side on the politics here. But this bipartisan
• 9:29
  panic shows that lawmakers across the board are
• 9:33
  realizing the regulatory framework is miles behind
• 9:37
  the actual capability of the technology. Absolutely.
• 9:40
  So if threat actors and enterprise defenders
• 9:42
  are spinning up thousands of these autonomous
• 9:45
  AI agents per minute, the physical constraints
• 9:48
  of the real world come into play. Someone has
• 9:51
  to power that compute. Right. The hardware. You
• 9:53
  would think the companies building the fiscal
• 9:54
  chips and the hardware infrastructure would be
• 9:57
  invincible right now. But Wall Street is telling
• 10:00
  a completely contradictory story. Let's look
• 10:02
  at the hardware paradox. The demand signal is
• 10:05
  undeniably astronomical. TSMC, you know, the
• 10:08
  massive foundry that actually manufactures the
• 10:10
  chips, had a record July. We are talking 14 .5
• 10:13
  billion U .S. dollars, up almost 45 percent year
• 10:16
  on year. And Intel saw so much demand for a share
• 10:19
  offering that they upsized it from $15 billion
• 10:21
  to $20 billion in a matter of hours. But here's
• 10:24
  where it gets wild. Look at Cisco and Applied
• 10:26
  Materials. Yeah. Cisco beat their expectations,
• 10:30
  pulling in $17 .3 billion in revenue, heavily
• 10:34
  driven by massive hyperscaler AI orders. But
• 10:38
  when they announced that... Their shares fell
• 10:40
  4 .6%. And Applied Materials set an all -time
• 10:43
  record with $9 .12 billion in revenue, and their
• 10:46
  shares fell 5%. Make that make sense for me.
• 10:49
  They're literally breaking records and printing
• 10:51
  money, and Wall Street punishes them. It comes
• 10:53
  down to two major factors, the concept of being
• 10:56
  priced for perfection and the reality of the
• 10:58
  infrastructure lag. Okay, let's start with priced
• 11:00
  for perfection. Wall Street has baked in such
• 11:03
  astronomical, compounding expectations for AI
• 11:05
  growth that merely having a record -breaking
• 11:08
  quarter is viewed as a failure. The market's
• 11:10
  tolerance for anything less than pure exponential
• 11:13
  acceleration is zero. And what about that infrastructure
• 11:15
  lag? Why is TSMC rewarded while Cisco is punished?
• 11:19
  Because building a data center happens in phases.
• 11:22
  The hyperscalers, Google, Meta, Microsoft, they
• 11:25
  buy the GPUs first. So TSMC, who fabricates the
• 11:28
  silicon, sees the immediate revenue spike. Okay,
• 11:30
  that makes sense. But you don't immediately buy
• 11:33
  all the networking gear to connect those chips
• 11:35
  until the facility is actually being built out.
• 11:38
  Cisco operates in that networking layer, which
• 11:40
  naturally lags behind the initial silicon purchase.
• 11:43
  Oh, I see. Furthermore, Applied Materials makes
• 11:46
  the massive complex machines that fabricate the
• 11:49
  chips. They sell to TSMC. Their peak happened
• 11:53
  even earlier in the cycle when the foundries
• 11:55
  were expanding. So Wall Street looks at Cisco
• 11:57
  and applied materials and assumes their hyper
• 12:00
  growth phase has already peaked, even if their
• 12:02
  current revenue is at an all time high. That
• 12:05
  context is crucial. So if you're a listener running
• 12:08
  IT for a midsize company, why do you care if
• 12:11
  Wall Street is throwing a tantrum over a stock
• 12:12
  price? Because financial volatility eventually
• 12:15
  translates directly into supply chain volatility.
• 12:18
  When these hardware companies get punished by
• 12:20
  the market. Despite record revenues, it impacts
• 12:23
  their internal capital expenditure plans, their
• 12:25
  hiring, their factory expansions. They might
• 12:28
  delay building a new networking equipment factory?
• 12:31
  Exactly. For the listener, this Wall Street drama
• 12:33
  eventually shows up on your desk as completely
• 12:35
  unpredictable lead times for your own company's
• 12:38
  router and server hardware refreshes. So you
• 12:41
  have a hardware market defined by volatile lead
• 12:44
  times and AI models moving at breakneck speed
• 12:47
  while being actively weaponized. How on earth
• 12:50
  is a standard, heavily regulated business supposed
• 12:53
  to safely implement AI? under these conditions?
• 12:56
  Well, they don't do it alone anymore. We are
• 12:58
  seeing a massive enterprise reality check. Look
• 13:01
  at the strategic partnership announced between
• 13:03
  IBM and OpenAI. They are embedding GPT 5 .6 into
• 13:08
  IBM Consulting Advantage and standing up a practice
• 13:10
  of thousands of consultants. Meaning the days
• 13:12
  of an enterprise just buying an API key off a
• 13:15
  website and handing it to their developers are
• 13:16
  ending? Exactly. For regulated industries like
• 13:19
  finance, healthcare, and government, systems
• 13:21
  integrators like IBM are the new gatekeepers.
• 13:25
  distribution route for frontier AI into the enterprise
• 13:27
  is now through a massive consulting firm with
• 13:29
  a certified practice that can handle the liability
• 13:31
  and the implementation strategy. And the rules
• 13:34
  governing that implementation are getting incredibly
• 13:37
  complex. Google Gemini just passed 1 billion
• 13:40
  monthly users. Crucially, 63 % of those interactions
• 13:44
  are voice and 20 % involve live camera or screen
• 13:47
  sharing. Which is a massive shift in how people
• 13:49
  use AI. Right. Meanwhile, Antropic is now watermarking
• 13:52
  all claw -generated text worldwide using a standard
• 13:56
  called C2PA, and they are doing this to comply
• 13:58
  with Article 50 of the EU AI Act, which went
• 14:01
  into effect on August 2nd. I want to break down
• 14:04
  how C2PA actually works because it isn't just
• 14:07
  a visible logo stamped at the bottom of a document.
• 14:10
  Right. It's way more technical than that. C2PA
• 14:11
  relies on cryptographic asset hashing. It mathematically
• 14:15
  binds the origin data, proving the content was
• 14:18
  generated by an AI directly into the metadata
• 14:21
  of the file itself in a way that travels securely
• 14:24
  with the asset, even if it gets copied or moved.
• 14:26
  Which brings up a fascinating geographic friction
• 14:28
  point. If I'm a listener running an IT department
• 14:31
  in Chicago, managing an industry, internal HR
• 14:33
  bot for my employees, why am I suddenly subject
• 14:37
  to the EUAI Act? You are experiencing regulatory
• 14:40
  export. For a global vendor like Impropic, it
• 14:44
  is a logistical nightmare to maintain one set
• 14:46
  of untracked models for the United States and
• 14:49
  a completely separate, cryptographically watermarked
• 14:52
  set of models for Europe. So they just do it
• 14:54
  for everyone. It is simply cheaper and more efficient
• 14:57
  to build one system that complies with the strictest
• 15:00
  global standard, in this case, Brussels, and
• 15:02
  apply it everywhere. So that mid -level IT manager
• 15:06
  in Chicago now has to ensure their internal HR
• 15:08
  systems can handle and process C2PA cryptographically
• 15:11
  hashed data simply because the base model they
• 15:14
  are renting complies with European law. The regulation
• 15:17
  exports itself globally through the software
• 15:19
  architecture. That is wild. And let's revisit
• 15:22
  those live camera stats for Gemini. If 20 % of
• 15:25
  a billion users are interacting visually, That
• 15:27
  completely shatters how most companies handle
• 15:29
  data security. If your company's acceptable use
• 15:32
  policy for AI only covers what employees are
• 15:35
  allowed to type into a text box, that policy
• 15:37
  is already obsolete. You have employees pointing
• 15:40
  their phone cameras at proprietary engineering
• 15:42
  schematics or confidential whiteboard meetings
• 15:44
  and asking the AI to analyze them in real time.
• 15:48
  It represents a massive, unmonitored expansion
• 15:50
  of the corporate attack surface. Which brings
• 15:53
  us to the immediate, boots -on -the -ground reality
• 15:55
  for security teams. While executives are busy
• 15:58
  strategizing over AI consulting contracts and
• 16:00
  EU compliance, the traditional IT house is currently
• 16:04
  on fire. Yeah, let's talk about that. Let's look
• 16:06
  at the fallout from a frankly brutal patch Tuesday,
• 16:09
  which brought around 400 fixes. One of them really
• 16:12
  stands out. CVE -2026 -6820 -0. It's a flaw in
• 16:17
  a Windows component called WinSOC. Microsoft
• 16:19
  rated it a CDSS 7 .0, categorizing it as important.
• 16:24
  Let's define CVSS for a moment. That is the common
• 16:26
  vulnerability scoring system. A 7 .0 out of 10
• 16:28
  usually means a vulnerability is serious, but
• 16:30
  perhaps requires the attacker to already have
• 16:32
  some level of access or it's difficult to execute.
• 16:34
  Right. So an IT team might logically schedule
• 16:37
  that patch for the following week's maintenance
• 16:39
  window. But the very next day, the security firm
• 16:42
  Checkpoint revealed that this exact vulnerability
• 16:44
  was already being actively exploited in the wild.
• 16:47
  North Korea's Lazarus Group was using it to deliver
• 16:50
  a kernel rootkit in an espionage campaign called
• 16:53
  Operation Dream Job. And a kernel rootkit is
• 16:57
  a nightmare scenario. It really is. It means
• 17:00
  the attackers have burrowed so deeply into the
• 17:02
  core of the operating system, the kernel, that
• 17:05
  they essentially have God mode access. They can
• 17:07
  hide their tracks from anti - So if a patch is
• 17:14
  rated a moderate 7 .0 on Tuesday, but on Wednesday
• 17:17
  we find out it's actively being used to deploy
• 17:19
  a North Korean kernel rootkit, are these CVSS
• 17:22
  severity scores? Basically useless as a prioritization
• 17:25
  metric? Well, it proves that a severity score
• 17:28
  is merely an input. It cannot be your only decision
• 17:30
  metric. The context of active exploitation is
• 17:33
  what matters. And the window of time you have
• 17:35
  to apply these patches is shrinking drastically.
• 17:38
  Just how fast are we talking? The data shows
• 17:40
  a VMware vCenter vulnerability that was exploited
• 17:43
  in just five days. An SAP Commerce cloud flaw
• 17:46
  was exploited in three days. Three days from
• 17:49
  the moment the patch was released to active in
• 17:51
  the wild attacks. And threat actors aren't just
• 17:53
  going after the obvious, heavily fortified corporate
• 17:56
  servers anymore. They are attacking the blind
• 17:58
  spots. Look at the breach at Siva Logistics,
• 18:02
  a global freight forwarder. The attackers breached
• 18:05
  Siva, but the fallout cascaded down the supply
• 18:07
  chain to six downstream brands, exposing customer
• 18:11
  data for companies like Valve's Steam Gaming
• 18:13
  Platform and ING Bank. Because supply chains
• 18:16
  are deeply interconnected data webs, your security
• 18:19
  is only as strong as your weakest vendor. I take
• 18:22
  it a step further because it goes beyond data.
• 18:24
  Look at the ransomware attack on the hospital
• 18:26
  in Winnipeg. The attackers didn't take down the
• 18:28
  clinical IT systems or the patient record databases.
• 18:31
  They took out the heating, ventilation, and air
• 18:34
  conditioning, the HVAC systems, and the physical
• 18:36
  door access controls. This is a critical vulnerability
• 18:39
  for so many organizations. If you ask a CSO to
• 18:43
  list their critical assets, they pull up a list
• 18:45
  of servers, databases, and employee laptops.
• 18:49
  They aren't thinking about the internet -connected
• 18:51
  thermostat in the lobby. Threat actors are aggressively
• 18:55
  targeting building management systems because
• 18:57
  they often aren't even on the IT department's
• 18:59
  asset register. If you aren't patching the HVAC
• 19:02
  system, a threat actor will gladly use it as
• 19:05
  a backdoor into your network. The ecosystem is
• 19:08
  incredibly fragile right now. We even saw a security
• 19:11
  researcher going by the handle Nightmare Eclipse
• 19:13
  publish a fully working zero -day exploit for
• 19:17
  Windows Defender. They called it Shield Break,
• 19:19
  and they released it simply because Microsoft
• 19:21
  threatened legal action over their previous vulnerability
• 19:24
  submissions. And there's no patch for it yet.
• 19:26
  Exactly. Speaking of immediate vulnerabilities,
• 19:29
  there is one highly practical, critical takeaway
• 19:31
  for our listeners today regarding Apple screen
• 19:34
  sharing. Yes. There is a critical flaw exposing
• 19:37
  Mac computers via port 5900. Attackers are scanning
• 19:41
  the internet, finding that open port, gaining
• 19:44
  root access to the machine, and silently installing
• 19:47
  neuro -cryptocurrency miners. Wow. If you have
• 19:50
  Macs in your fleet with port 5900 exposed to
• 19:53
  the open internet, closing that port needs to
• 19:55
  be your top priority the moment this deep dive
• 19:57
  ends. Let's take a breath and recap the sheer
• 20:00
  scale of the terrain we've covered today. We
• 20:03
  started by unpacking how half a trillion dollar
• 20:05
  AI infrastructure deals are now entirely reliant
• 20:08
  on debt financing, creating a massive supply
• 20:11
  surge that is plummeting the cost of AI models.
• 20:13
  Huge shift. We saw how that incredibly cheap
• 20:16
  compute is directly empowering threat actors
• 20:18
  to run near autonomous hacking swarms. We navigated
• 20:21
  the volatile hardware market, where the infrastructure
• 20:23
  lag and impossible expectations are punishing
• 20:25
  companies despite record profits. We looked at
• 20:28
  how global regulations like the EU AI Act are
• 20:30
  exporting themselves into your local systems
• 20:32
  via cryptographic watermarks. And we ended with
• 20:35
  the stark reality of three -day exploit windows,
• 20:38
  compromised hospital HVAC systems, and North
• 20:41
  Korean rootkits. It is a phenomenal amount of
• 20:44
  structural change occurring all at once. It really
• 20:47
  is. And it leaves me with one final thought tomorrow.
• 20:49
  over building on everything we've discussed today
• 20:51
  yeah we are watching ai rapidly drive down the
• 20:54
  cost and complexity of offensive cyber operations
• 20:57
  you know those open source agent swarms running
• 21:00
  autonomous learning cycles but simultaneously
• 21:02
  we are seeing ai deployed for defense like open
• 21:06
  ai's daybreak models actively hunting for zero
• 21:09
  -day vulnerabilities right it begs a serious
• 21:11
  question are we rapidly approaching a point where
• 21:13
  human security teams are no longer the primary
• 21:15
  combatants are we simply going to be spectators
• 21:18
  watching offensive AI, battle machine speed,
• 21:21
  defensive AI in real time, completely unable
• 21:24
  to intervene because the operational tempo is
• 21:26
  entirely beyond human cognition. Wow. That is
• 21:29
  a heavy but incredibly necessary question to
• 21:31
  sit with as we watch these tectonic plates shift
• 21:33
  in real time. The map has changed, and it is
• 21:36
  entirely up to us to learn how to navigate the
• 21:39
  new terrain. Thank you so much for joining us
• 21:41
  for this deep dive. As a casual reminder before
• 21:43
  you go, check those vendor contracts to see if
• 21:45
  you can get a better rate on your AI compute,
• 21:47
  update your acceptable use policies to account
• 21:49
  for live cameras, and for goodness sakes, close
• 21:52
  port 5900 on your Macs. We'll catch you on the
• 21:54
  next one.