Last Week in Tech – Google Reshuffles Its AI Empire, Washington Writes a Rulebook It Won't Show You, and Memory Sold Out to 2030
24 minAudio
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! Your Monday roundup of the technology and security news that matters to Australian businesses, covering 3-9 August 2026. In this episode: Google's AI leadership shake-up (5 Aug). Demis Hassabis moves to Chair of Google DeepMind and Chief Scientist of Alphabet. Koray Kavukcuoglu becomes SVP of Google DeepMind reporting to Sundar Pichai, taking Gemini research and product. Jeff Dean leaves after 27 years to co-found a public benefit corporation with Sanjay Ghemawat, with Google as founding investor and cloud partner. The Gemini app is past 950 million monthly users; Gemma is past 900 million downloads. Washington's unpublished frontier-AI framework (4 Aug). Under Executive Order 14409, an NSA-led group has built a process for up to 30 days of pre-release government access to frontier models - and reportedly will not publish it. Open-weight models are excluded. Why a shrinking public evidence base means you need your own evaluation set. AMD and Palantir (3-4 Aug). AMD: revenue $11.5bn (+50%), Data Centre $6.7bn (+107%) and now 58% of the company, Q3 guided to about $13bn. Palantir: revenue $1.935bn (+93%), US commercial +149%, adjusted free cash flow $1.22bn, guidance raised to about $8.15bn. Two great quarters, two opposite market reactions, and what that says about capital intensity. Microsoft's 10-K discloses $24.1bn of OpenAI revenue, plus $6bn of accounts receivable from a single customer as at 30 June. A prompt to run your own customer-concentration numbers. Memory is being sold out to 2030. Samsung plans to cover 60-70% of its memory production capacity under five-year binding agreements with advance payments and floor prices; SK hynix has completed talks with around ten customers. What allocation-constrained supply means for your FY27 hardware budget. Meta launches Muse Code (5 Aug), and Bending Spoons agrees to buy Airtable for a $1.285bn enterprise value (4 Aug). Security round-up: the Metabase zero-day and the Framework, Tally and LexisNexis fallout; IBM Langflow and Apache Tomcat added to CISA KEV; and a preview of Friday's N-able N-central deep-dive. Visit www.kinsoft.com.au to talk through your security and IT needs. Sources: Google blog, "The next chapter of our AI momentum" (5 Aug 2026); Axios and Tech Policy Press (4 Aug 2026); White House Executive Order 14409; AMD Q2 2026 results (4 Aug 2026); Palantir Q2 2026 results (3 Aug 2026); Microsoft FY2026 Form 10-K; The Korea Herald and StorageNewsletter (2-3 Aug 2026); TechCrunch, CNBC and the Wall Street Journal (5 Aug 2026); Bending Spoons investor newsroom (4 Aug 2026); BleepingComputer, TechCrunch and The Hacker News (6-8 Aug 2026); CISA Known Exploited Vulnerabilities catalogue (3-4 Aug 2026); Huntress (1-7 Aug 2026).
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
I mean, imagine discovering that the richest company on earth has six billion dollars tied up in an IOU from just one single customer. Right. Just sitting there. Yeah. Just sitting in accounts receivable. Six billion dollars. And it's basically this massive single point of failure buried deep in a, you know, just a totally routine financial disclosure. It really is wild when you catch it. It is. And so today we're taking the crisp bow off the tech industry's shiny PR to actually look at the invisible wires and, well, the massive
debts that are holding it all together. Because it changes your entire perspective, right? When you stop looking at the flashy product launches and start looking to hit independencies. Exactly. We all want to believe that the software and the hardware ecosystems that we rely on are these robust, independent pillars. But a closer read of the data shows a deeply intertwined system. One pulled thread can just unravel your entire strategic roadmap. And that is exactly our mission for this deep dive. We are unpacking a highly dense tech briefing from August 17th, 2026. The
Tech Talks with Kinsoft Update. Right, the Kinsoft Update. We are going to decode a series of quiet but massive tectonic shifts happening in the background of the industry right now. So we're looking at a major changing in the guard at Google. classified government AI testing frameworks, those hidden billion dollar financial dependencies we mentioned, and a physical hardware squeeze that is literally going to dictate what you can actually buy next year. And if you're looking for the through line for all of these seemingly
disconnected events. It's control. Control. It's all about who actually builds the foundational AI, who is legally allowed to evaluate it, who's financing all this compute power, and frankly, who holds the keys to the physical supply chain. Because if you don't understand those control mechanisms, you really can't protect your own operations. Exactly. So let's start at the absolute top of the AI food chain with who is actually building the tools you use every day. Because Google just went through an earthquake. Oh, Matt. Yeah. On August 5th, they dropped two announcements
within like... minutes of each other, and it detailed the biggest AI leadership reshuffle in years. So, Demis Hassabis, the legend who co -founded DeepMind, is stepping back from day -to -day operations. Right. He's moving up. Yeah. He's becoming the chair of Google DeepMind and Alphabet's chief scientist. And taking his place operationally is Kare Kavukoglu. And to be clear, he is no outsider parachuting in to just, you know, manage spreadsheets. Definitely not. He is a 13 -year veteran of DeepMind. I mean, he
worked on foundational projects like DQN. Oh, DQN. That was the reinforcement learning that famously taught AI to play the classic Atari games. That's the one. He also worked on WaveNet, which is... the architecture behind that incredibly realistic AI voice generation. And so now he steps up as senior vice president reporting directly to Google CEO Sundar Pichai. Which is huge. It is. He now controls Gemini model development, the frontier research teams, the developer teams, and the Gemini app itself. And we should note the briefing highlights the Gemini app just crossed
950 million monthly users. It's a staggering number. I know, right? But before we get into the implications of Coray taking over. The second bombshell in that same announcement was about Jeff Dean. 27 years at Google. 27 years. He's the mind behind TensorFlow, Google Brain, MapReduce. These are technologies that literally shaped the modern Internet. Completely. And he is leaving. He's starting an independent public benefit corporation with his longtime collaborator, Sanjay Gemmawat,
focusing on machine learning and science. Now, Google is a founding investor, but still. This is the end of a massive era. It is a profound changing of the guard. But, you know, for the listener who is actually relying on Google Workspace or Google Cloud for their business, the structural shift with Coray is what actually matters to your day to day. How so? Well, historically, Google suffered from this massive paper to product gap. Oh, right. The brilliant researchers publishing cutting edge AI papers sat in a completely different
silo from the product teams trying to build features for everyday users. That separation created massive internal friction. So if Google was like a massive world -renowned restaurant, it sounds like the visionary executive chef who spent all day experimenting with molecular gastronomy just moved up to the corporate office. I like that. Yeah. And the veteran sous chef who actually knows how the kitchen line operates and how to get plates to tables quickly is now entirely in charge of the menu. That is exactly the dynamic at play. Consolidating
the cutting edge research and the consumer product pipelines under one operational leader and having them report straight to the CEO signals a clear mandate for speed. They want to ship faster. Exactly. The pace of AI features rolling out to your workspace is going to accelerate. But, and this is a big but, this introduces a critical governance check for anyone managing enterprise tech. Right. When a hyperscaler, one of these massive cloud providers, changes the singular person deciding what gets built, your own internal roadmap assumptions suddenly have a new owner.
So if your business heavily relies on Google's AI trajectory, you need to be asking questions in your next account review about how this specific leadership change alters their promised deliverables. You have to. You can't just assume the old roadmap is still valid. Okay, so Google is trying to speed up its product pipeline and push AI out faster. But here is where I get genuinely confused by the source material. Because while big tech is hitting the gas, the US government... is quietly
like pulling the curtains closed on how these models are even evaluated. Yeah, this is a major shift. The briefing details Executive Order 14409. The government has directed an interagency group led by the National Security Agency to build a framework for reviewing what they call frontier AI models. And the mechanism of this framework is severe. It requires developers of state -of -the -art closed source models to give the government pre -release access for up to 30 days before they can even ship the product to the public. Wait, I need to stop you there. A 30 -day secret
testing window. Yep. How does the government even define what a national security risk is in a text generating model? The briefing itself points out there is no clear definition of state of the art yet. And logistically, how do you comprehensively test a massive, unpredictable neural network in just a month? You are hitting on the exact logistical nightmare the industry is facing right now. They're essentially putting these models in a digital SCIF. A sensitive compartmented information facility. Right. Highly secure, physically
isolated environments with incredibly detailed access logging. They will use automated red teaming, bombarding the AI with prompts designed to trick it into generating, say, biological weapon blueprints or discovering zero day software vulnerabilities. That sounds intense. What about the open models, though? Interestingly, open weight models like Google's Gemma, which just hit 900 million downloads, they are explicitly excluded from this pre -release framework, mostly because once an open model is out, the cat is out of the bag. But for the
massive proprietary models from OpenAI, Anthropic, or Google, this testing is mandatory. And the most concerning detail in the briefing, at least to me, is that the government is reportedly not going to publish this framework or the results. No, it's entirely classified. They've actually directed the National Institute of Standards and Technology NIST to stop issuing public model assessment reports. So if they're testing these models in secret and shutting down the public scorecards, how on earth... Are businesses supposed to know if the AI they are integrating is actually
safe? That's the problem. The public evidence base is effectively vanishing. For the last two years, if you wanted to know whether a new model was prone to leaking data or writing insecure code, you could just read an independent public evaluation from NIST. Right. With the cyber capability benchmarking process becoming classified, you are left with nothing but the vendor's own marketing press releases. And a vendor is never going to tell you their new upgrade is worse than the last one. So the burden completely shifts to the listener. If I'm running a company, I can't rely on generic public benchmarks anymore to
decide if an AI tool is safe for my customer data. You must build your own evaluation sets immediately. How does that look in practice? You need to curate 30 or 40 real... documented cases from your own business where you know the absolute correct answer every single time your vendor pushes a mandatory model upgrade you run that specific set of questions through the new system that makes sense it takes maybe a half day of work to build that baseline but it is the only empirical way to know if an upgrade you didn't even ask for just quietly broke your
internal processes Man. Well, if the government is no longer giving us the public safety metrics on AI, the next logical place to look is how the financial markets are evaluating AI's success. Oh, absolutely. And looking at the August earnings reports, the market is looking aggressively under the hood at capital intensity. The reaction to these earnings calls was wild. Look at AMD. Yeah, that was a fascinating one. They reported $11 .5 billion in revenue, up 50 % year over year. Data center revenue was up 107 percent, now making up 58 percent of the entire company. Lisa Su
confirmed a massive partnership to deploy GPUs and AMD Helios racks for Anthropic. I mean, it's an incredibly strong quarter, but their shares fell. Yeah, they dropped. And compare that to Palantir, who reported the day before. Palantir posted just under $2 billion in revenue, up 93 percent. But the metric that mattered was their adjusted free cash flow, which came in at a massive $1 .22 billion. And following that report, Palantir's shares jumped sharply. They did. I have to challenge
the market logic here. Why punish AMD for growing 50 %? AMD is mining the iron and forging the steel that the entire AI industry requires to function. Palantir's just selling tickets to use the bridge. Why does the market reward the software layer but penalize the company laying the physical foundation? Because laying that physical foundation requires an eye -watering amount of upfront cash, and investors are getting nervous about the long -term math. The key metric to look at is capital expenditure versus gross margin. Okay, break that down. AMD's guidance
showed their gross margin staying roughly flat around 56%, while the money they had to spend on capital expenditure, so building new fabrication capacity and research, that more than doubled. Palantir, being a software company, threw off $1 .2 billion in free cash flow on $1 .9 billion of revenue. Ah, I see. The market isn't just counting the revenue AI generates anymore. They're aggressively questioning what it physically costs to earn that revenue. Exactly. And this reveals a structural lesson about the economics of AI. You have to understand the difference between
an AI product and an automated service. What's the difference? If you build a true software product, you spend the money once to develop it, and your margins expand as millions of people use it. But with many AI workloads, the computing cost scales one for one with every single user prompt. Because it takes GPU power every time someone asks a question. Right. That isn't a product. That is a service you happen to have automated. And if you don't price it carefully, scaling up your user base will literally bankrupt you in compute costs. Which brings us perfectly
to the most jaw -dropping financial vulnerability in this entire briefing. Those massive compute costs have to be paid somehow. And it leads to some terrifying accounting. Let's look at Microsoft's fiscal 2026 Form 10 -K disclosure. This is where it gets scary. It really is. Microsoft disclosed $24 .1 billion of revenue from commercial arrangements with OpenAI for the year ending June 30. That number alone is staggering. But the real shocker is further down the page. Microsoft disclosed
$6 billion in accounts receivable from OpenAI. Accounts receivable is money that you have billed for services rendered, but the client has not actually paid you yet. Six billion dollars owed by one single customer explicitly disclosed in a statutory filing. So if I'm running a midsize business and I see the richest tech company on earth holding a six billion dollar IOU from a single partner, I mean, that tells me I need to look at my own customer concentration risk immediately. That is the exact right takeaway. Customer concentration is a blind spot that companies systematically under report to themselves. Oh,
yeah. Many businesses could not tell you exactly what percentage of their operating revenue is tied up in their top three clients. Ask yourself, what happens to your runway if your biggest client suddenly delays their invoices by 90 days? If customer concentration is a critical metric worth disclosing to the SEC at Microsoft scale, you absolutely must be tracking it at yours. Wow. Well, to cover those massive capital expenditures and unpaid IOUs, companies have to guarantee their future capacity to generate revenue. And that brings us to the physical hardware market.
The squeeze. Yeah, the great hardware lockup, because to guarantee that capacity, the megacorporations are quietly buying the supply chain out from under everyone else. Let's look at memory components. Samsung, the world's largest memory manufacturer, is putting 60 to 70 percent of its entire memory production capacity under five year binding agreements. That's a massive chunk of global supply. And these contracts are going to just five major global data center customers and five massive AI clients. Deposits are already paid. And crucially,
floor prices are set to protect Samsung if market prices fall. SK Hynix is executing the exact same playbook with around 10 of their own customers. The structural implications of this are massive. I mean, AI doesn't just require processing GPUs. It requires staggering amounts of high bandwidth memory to feed data to those processors. By signing these binding five -year agreements, the vast majority of the global memory supply for the rest of the decade is contractually locked behind closed doors. It feels like trying to buy a laptop
or a server for your IT team is going to be like trying to buy front row concert tickets, only to find out the venue sold 70 % of the seats directly to a handful of VIPs before the box office even opened. Yeah, that's exactly it. The rest of us are left fighting over whatever scraps are left in the nosebleeds. That is painfully accurate. The standard enterprise buyer is not one of those VIPs. Neither is your primary IT reseller. The memory chips going into your next server cluster or the laptops you plan to issue to employees over the next three years are being
pulled from a much smaller, highly constrained leftover pool. And what about pricing? Well, because of those floor price clauses, the historical boom and bust cycle of memory pricing is broken. Even if global consumer demand softens, the manufacturers have a contractual safety net, meaning prices are unlikely to drop for the everyday buyer. So how do we actually play defense against a global hardware squeeze? You remove ambiguity from your procurement pipeline immediately. Go to your suppliers today and get your fiscal year
2027 hardware quotes in writing with a locked validity period. a verbal estimate is worthless when allocation gets tight right get it in writing second if you have a critical infrastructure refresh scheduled for the next 12 to 18 months you need to model the financial impact of bringing that purchase forward waiting will likely subject you to higher prices and massive delays ask your supplier explicitly Is this specific configuration currently allocation constrained? Okay, so we are seeing scarcity in hardware, but we are also
seeing major turbulence in software, specifically the software as a service market. The briefing highlights Airtable being acquired by Bending Spoons for an enterprise value of $1 .285 billion. Quite a drop. Yeah. To put that in perspective, Airtable serves over 500 ,000 organizations, including 80 % of the Fortune 100, and was valued at over $11 billion back at its peak in 2021. This represents a brutal market recalibration for SaaS valuations. But what matters to the listener who relies on Airtable is understanding
the acquirer's standard playbook. Right, bending spoons. Yeah, bending spoons is a private equity style firm that also owns Evernote, Vimeo, WeTransfer, and recently picked up AOL and Eventbrite. pattern after an acquisition is to radically reorganize the internal teams and aggressively overhaul pricing and monetization structures to extract value. Right. So if Airtable is a load -bearing pillar for your company's daily operations, this is the exact moment to pull your contract, read
the renewal terms regarding price hikes, and make sure you know exactly how to export your data if the new pricing model doesn't work for you. You need an exit strategy before you are forced to need one. Exactly. And speaking of shifting software dependencies, we have to look at the tools developers use to actually write software. Meta just launched a beta for Muse Code, powered by their Muse Spark model. Which is their new coding agent. Right. This is a terminal -based AI coding agent designed to compete directly with Anthropic's Claude code and OpenAI's Codex.
The technical differentiator Mark Zuckerberg highlighted is its use of isolated work trees. Okay. So how does an isolated work tree actually function in practice? Essentially, instead of the AI fumbling around in your live code base, it fans out and creates separate sandboxed clones of your repository. It assigns sub -agents to work on different problems in parallel within those isolated environments. So it's not touching the real code yet. Exactly. Only when the code is verified does it present the solution back
to your main working copy. It is designed to prevent the AI from accidentally breaking production code while it hallucinates a solution. The source material did mention some industry reporting about pricing structures from used code, including a rumor regarding a cheaper tier that requires you to consent to your proprietary code being used to train their future models. But the source could not verify those details, so we want to be very clear that the pricing rumor remains unverified. Good to clarify. Still, the broader trend is undeniable. Massive tech companies are fiercely competing to provide agents that write
your internal code. Which brings us to an urgent governance requirement. If you run a development team, you must establish a firm written policy today. Which of these AI agents are legally permitted to touch your company's code repositories? And under what specific conditions, if any, can your code be utilized for external training? Because if a developer on your team uses an unauthorized agent to fix a bug and that agent absorbs your proprietary algorithm, you have a massive intellectual property leak. You do. And it happens faster
than you think. All right. We've covered long -term AI strategy, financial risks, and hardware scarcity. But we always need to close out with the immediate fires. The things you have to fix today. Yes, the things you literally have to fix today because the structural dependencies are already failing. The urgent patch list for August is severe. On August 3rd, a zero -day vulnerability was exploited in Metabase, an incredibly popular open -source business intelligence tool. And this is a bad one. This is an unauthenticated SQL injection via the password reset endpoint. Because Metabase is an analytics tool, it is
by definition wired into the core databases of your entire company. For those who might not live in the cybersecurity trenches, how does an unauthenticated SQL injection actually work in this scenario? Imagine walking up to a bank teller. And instead of handing them your ID to access your account, you hand them a piece of paper with the bank's own internal override code that says, give this person all the cash in the vault. Oh, wow. Yeah. And SQL injection means the attacker types database commands directly into the password reset box. The system fails to recognize it as malicious input. reads the
command and simply obeys, granting full administrator access. That is catastrophic. The attacker instantly gets the credentials for every single database that the Metabase instance is connected to. The briefing notes that Framework, the laptop maker, had to notify their entire customer base. A form builder called Tally reported email addresses and hash passwords were taken. LexisNexis saw unusual activity on third -party hosted servers affecting their Metabase API. It was widespread. But here is the part that makes no sense to me.
The source says this vulnerability has no CVE score. The Common Vulnerabilities and Exposure System hasn't officially rated it, yet Metabase itself declared it a 10 out of 10 emergency. How does a 10 out of 10 not have an official CVE? It highlights a terrifying reality of modern cloud exportation. The speed of attackers actively exploiting a flaw often outpaces the bureaucracy of assigning a CVE number. You cannot wait for an official government severity score to act. A 10 out of 10 from the vendor means you drop
everything. So what's the fix? The fix requires you to patch to the latest secure August release immediately. You must manually revoke all active user sessions, rotate every single connected database credential because they must be considered compromised, and if you cannot patch today, you must block the password reset endpoint at your network perimeter. Okay, drop everything and patch Metabase. What else is on the critical list? And Abel's nCentral platform. This is a remote monitoring and management tool used by countless outsourced IT providers. It has been under active exploitation since August 1st due
to an authentication bypass. An authentication bypass. So they just skipped the log. Yep. Attackers can simply step around the login screen and gain full administrative control of the IT management systems. If you are running this on -premises, the hotfix build released on August 6th is absolutely mandatory, even if you already applied an earlier patch. So patch Metabase and nCentral today. Also, briefly, IBM Langflow and Apache Tomcat were just added to Cease's known exploited vulnerabilities catalog, so ensure your teams are patching those
on their required timelines as well. You know, these vulnerabilities tie perfectly into our broader theme today. Metabase is your analytics plane. And central is your management plane. When attackers target these specific tools, they aren't just trying to break a single application. They're going after the foundation. Exactly. They are weaponizing your own structural dependencies against you. Having absolute visibility and rigid control over these central nervous systems is the only way to survive. It really all comes back to looking past the shiny box to see the
invisible wires. From Google re -architecting its AI leadership to eliminate internal friction, to the US government quietly locking down model testing and killing public benchmarks. We've traced $6 billion IOUs buried in Microsoft's regulatory filings, watched memory chips get contractually hoarded by megacorporations years in advance, and seen critical zero -day flaws threaten the very tools we use to manage our data. It proves that being genuinely well informed in the tech sector right now requires ignoring
the press releases. It means reading the boring statutory filings to find the hidden debt, understanding the capital expenditure required to keep AI running, and tracking the physical supply chains that determine if your software actually has a server to run on. Because if you only look at the final product, you are completely blind to the risks holding it up. It leaves me with a final, slightly provocative thought to consider. Let's hear it. We are seeing governments hide AI benchmark results from the public, and we are seeing big tech hide massive financial and hardware dependencies until
statutory filings force their hand. Given how brittle this ecosystem is, how long until the market demands a mandatory nutritional label for enterprise AI tools? Oh, a nutritional label. Yeah. Not just a label for algorithmic safety, but a transparent breakdown proving the fundamental economic and structural stability of the tools we are building our entire future upon. An AI nutritional label for structural stability. Yeah. Imagine seeing a label that says, warning. This tool is heavily dependent on a single, uncollected
$6 billion invoice. Exactly. It feels completely inevitable. Thank you for joining us on this deep dive. Keep questioning the information around you. Lock in your hardware quotes this week and start building those AI evaluation sets today. We'll catch you next time.
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! Japanese home goods company Iris Ohyama (online store operator Iris Plaza) found a suspicious file in an external...
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! Australian online retailer Who Gives A Crap told customers on 7 November 2026 that their information was accessed...
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! Last week in tech, 7 to 14 November 2026. Microsoft's November Patch Tuesday (10 November) fixes well over a...
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! Chainalysis disclosed on 30 October 2026 that an external actor using "offensive agentic techniques" accessed...
Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! NSW Health reported unauthorised access to its Oral Health Fee For Service Scheme website on 29 October 2026;...