Latest / Deep Values: Value Investing Deep Dives / ZUCKERBERG SLAMS! How Meta’s Internal Chaos Crashed AI Market
Transcript
- 0:00So on July 1st, 2026, there was this, the single phrase that
- 0:05leaked to the press and it basically wiped out billions of
- 0:08dollars in global wealth in just a few hours.
- 0:11Right, Yeah, the phrase was excess AI compute.
- 0:14Exactly. Excess AI compute.
- 0:16Which is crazy, because for over a year, the entire tech
- 0:21narrative was built on this one solid, unshakable foundation,
- 0:24Right? Like artificial intelligence,
- 0:26computing power was the most scarce, most valuable resource
- 0:30on the planet. Everyone was just desperate for
- 0:32more. Oh absolutely was a gold rush.
- 0:34And then suddenly Meta, you know, the biggest buyer in the
- 0:37world by far, just announces they're launching a platform to
- 0:40sell off their extras like they were full.
- 0:42They didn't eat it anymore. And I mean, if that wasn't bad
- 0:44enough, the very next day this leaked audio recording hits the
- 0:48Internet. It was Mark Zuckerberg at an
- 0:49internal all hands meeting, and he's just flatly telling his
- 0:52employees that AI agents were, well, they weren't accelerating
- 0:56as expected. Which is just a massive red
- 0:58flag. Huge.
- 0:59And the combination of those two events, it didn't just rattle
- 1:02the stock market. I mean, it triggered a global
- 1:04avalanche, the Philadelphia Semiconductor Index, the Sox, It
- 1:08plummeted over 6% right at the opening bell.
- 1:11Wow. Well, welcome to today's deep
- 1:12dive everyone. We are looking into one of the
- 1:14most fascinating and honestly destructive periods in the
- 1:18global AI market. Our sources today are two
- 1:21incredibly detailed analysis from the YouTube channel You
- 1:25Know or Best Partner, which breakdown the scathing report
- 1:28from Semi Analysis, as well as the market data from that
- 1:31massive July 2026 tech crash. Yeah, our mission today is
- 1:34really to uncover how Meta's mind boggling infrastructure
- 1:37spending spree inflated this global AI stock bubble, why that
- 1:41bubble violently popped, and honestly, the deeply hidden,
- 1:44shocking cultural dysfunctions inside Meta's multibillion
- 1:47dollar tech teams. Right, because we're talking
- 1:50about a scenario where a single company's internal missteps
- 1:53triggered a global chain reaction.
- 1:55And it reveals some really hard truths about the actual state of
- 1:58this whole AI race. But you know, before we jump
- 2:01into the absolute carnage of July 2026, since we're talking
- 2:05so much about market volatility today, a quick tip for you
- 2:08listening. All right, the trackers.
- 2:10Yeah, if you want to track total market temperature yourself
- 2:13without relying on the crazy hype cycle, welcome to search
- 2:16and download the iOS app Buffett Indicator from the Apple App
- 2:19Store or Bubble Watch from Google Play.
- 2:21And also to access more advanced researches, applications and our
- 2:25latest Deep Values agents. Welcome to explore and join the
- 2:28Deep Values APP. You can find that at the Apple
- 2:31App Store link in our show notes.
- 2:32It's apps.apple.com/us/apps slash Deep values and so on.
- 2:36Check the notes for the full link.
- 2:37Perfect. And I should probably just add
- 2:39our standard disclaimer here too, right?
- 2:41So all analysis is based on publicly available sources and
- 2:45may be subject to revisions or differing interpretations.
- 2:49All contents from this channel is for research and
- 2:51entertainment purposes only and does not constitute any
- 2:54investment or financial advice or service.
- 2:56Conduct your own due diligence before making investment or
- 2:59financial decisions. Awesome.
- 3:01OK, so let's get into the fallout of that leaked audio
- 3:04because the panic, it wasn't just contained to Wall Street.
- 3:08Like if you look at South. Korea.
- 3:09South Korea got hammered. Yeah, the.
- 3:10Kospi index suffered its worst single day drop since the 1998
- 3:14Asian financial crisis. They went down almost 11%.
- 3:18Heavyweights like Samsung and SK Hynix, their valuations just
- 3:22completely cratered. And the sources point out that
- 3:24this was particularly devastating for retail investors
- 3:27in South Korea. Right the the Ant investors as
- 3:30they call them. Exactly because they were
- 3:32heavily leveraged, over 400,000 retail accounts faced immediate
- 3:37margin calls. Yeah.
- 3:39And for anyone unfamiliar with the mechanics of a margin call,
- 3:42it is a brutal, brutal financial reality.
- 3:45Basically, these retail investors had borrowed money
- 3:47from their brokers to buy more tech stock, essentially betting
- 3:51that the AI boom would just, you know, go up forever.
- 3:53Because everyone said it would. Right.
- 3:56But when Meta's announcement sent the stock prices into a
- 3:58freefall, those brokers demanded immediate cash to cover the
- 4:01loans. And when the investors couldn't
- 4:04pay, the brokers automatically liquidated their accounts.
- 4:08I mean, hundreds of thousands of people had their savings wiped
- 4:10out in seconds. China's a share tech sector also
- 4:14tanked. It was just an indiscriminate
- 4:15sell off. The analogy that always comes to
- 4:17mind for me here is like, it's like someone running into a
- 4:19sports stadium, right? And they aggressively buy up
- 4:23every single hot dog, jersey, foam, finger, whatever at a
- 4:271000% markup. Just throwing cash everywhere.
- 4:30Yeah, and the vendors, they think, oh, wow, this is the new
- 4:33normal, we're rich. But then suddenly that guy blows
- 4:35a whistle, says actually I'm bored, Anyone want to buy my
- 4:38extras? And walks out the door.
- 4:40The entire local stadium economy just collapses.
- 4:44But the question is, how did the entire global market get so
- 4:47inextricably linked to just Meta's buying habits in the
- 4:50first place? Well, to understand that
- 4:52gravitational pull, we really have to trace the narrative back
- 4:55to like 2025 and early 2026. Meta push their AI capital
- 5:01expenditure, which is, you know, their budget for buying physical
- 5:04hardware and building data centers, to an astronomical 125
- 5:08billion to $145 billion. Which is just a number so big
- 5:11it's hard to even process. It really is.
- 5:13I mean that single company accounted for nearly one fifth
- 5:17of the entire United States tech sectors AI investment.
- 5:20They were single handedly keeping the lights on at
- 5:22multiple hardware manufacturers. Let's let's put that $145
- 5:27billion into perspective for a second.
- 5:29That is roughly the entire gross domestic product of a small
- 5:33country like Morocco, entirely dedicated to buying servers and
- 5:37chips. Wow.
- 5:39Yeah, that puts it in perspective.
- 5:40And it wasn't just hardware, right?
- 5:42They signed over $100 billion in external compute contracts, and
- 5:46they handed 60 billion of that to AMD alone.
- 5:50Massive, massive contracts. They paid $143 million for a
- 5:56mere 49% stake in Scale AI. And my favorite crazy stat?
- 6:01We're talking about $100 million signing bonuses handed to
- 6:04individual researchers from Open AI just to get them to walk
- 6:07across the street. Yeah.
- 6:08And if we connect this to the bigger picture, this brings us
- 6:11to a crucial economic mechanism called the anchoring effect.
- 6:15Mehta wasn't just, you know, participating in the AI market.
- 6:18By throwing around what appeared to be an infinite bottomless
- 6:21budget, they were setting a completely irrational price
- 6:24anchor for the rest of the world.
- 6:26Right, the anchor. Yeah, in behavioral economics,
- 6:30anchoring happens when people rely way too heavily on the
- 6:33first piece of information they receive.
- 6:35Like if you go to buy a house and the first one you tour is
- 6:39some mediocre tear down listed for $10 million, a $5,000,000
- 6:43house down the street suddenly feels like a bargain, even if
- 6:46it's really only worth 1,000,000.
- 6:48Oh, that makes total sense. So when venture capitalists and
- 6:51institutional investors were looking at other AI startups,
- 6:55their baseline for what was normal was completely warped by
- 6:59Mark Zuckerberg's checkbook. Exactly.
- 7:01Companies like Core, Weave, and Nebius, they saw their
- 7:04valuations skyrocket simply because they existed in the same
- 7:07ecosystem that Meadow was aggressively inflating.
- 7:10The market collectively assumed that, well, if the biggest, most
- 7:13data rich buyer in the world was willing to pay these exorbitant
- 7:15prices, they must know something we don't.
- 7:18This must be the new normal. But then the reality check
- 7:21arrived. We have to ask why Meta suddenly
- 7:24retreated from that massive anchor they created.
- 7:26And the underlying issue, according to the sources, was
- 7:29that their actual product, the Lama 4 model, severely
- 7:32underperformed expectations. It really did.
- 7:35When Llama 4 was released, it faced widespread criticism from
- 7:39the developer community. I mean, there were serious
- 7:41allegations that the model was essentially cheating on
- 7:44benchmark tests. And just to clarify, cheating on
- 7:47benchmarks and AI basically means the model was over
- 7:49optimized for the specific data sets used to grade it, but it
- 7:53lacked actual generalized reasoning skills.
- 7:56Right, it's not actually smart. Exactly, it's the equivalent of
- 8:00a student memorizing the answer key to a standardized test
- 8:03without actually learning the underlying math.
- 8:06Sure they score 100%, but if you ask them to solve a real world
- 8:09problem they just freeze. In practical application.
- 8:12LAMA 4 was lagging noticeably behind Frontier models from Open
- 8:16AI, Anthropic and even Peep Seek.
- 8:19So Meadow was spending the most money by a massive mulching, but
- 8:22they were not producing the best model.
- 8:24And if we look at the historical context, this is actually a
- 8:27recurring behavioral pattern for Meta's leadership.
- 8:29Massive, flashy spending sprees are often a smokescreen for
- 8:32internal panic. Oh, like the metaverse pivot?
- 8:35It's exactly In 2021, Mark Zuckerberg famously renamed the
- 8:39entire company to Meta and burned over $70 billion on
- 8:42Reality Labs to build the metaverse.
- 8:45Right, and that metaverse pivot was a direct reaction to missing
- 8:48the transition to native mobile apps and getting completely
- 8:51outflanked by TikTok in the short video era.
- 8:54Like Horizon Worlds, their flagship VR product, it barely
- 8:58had any users compared to gaming platforms like Roblox, yet they
- 9:02were spending 10s of billions on it.
- 9:03Yeah, and this AI spending spree was the exact same playbook.
- 9:06It wasn't a calculated strategic masterstroke.
- 9:09It was a panic buy. Duckerbert was just terrified of
- 9:12missing the next technological paradigm shift, so he decided to
- 9:15buy every available piece of hardware to ensure he had a seat
- 9:17at the table, regardless of whether his teams actually knew
- 9:20what to do with it. OK, let's pause here and unpack
- 9:23this a bit. I want to push back on this
- 9:24narrative because even if Llama 4 wasn't great right out of the
- 9:29gate, Meta is legendary for its engineering talent.
- 9:33I mean their motto used to be move fast and break things.
- 9:36Surely possessing all this custom hardware, securing all
- 9:39these chips and having total control over their supply chain
- 9:42gives their internal engineers a massive structural advantage to
- 9:46just iterate and build Llama 5 right?
- 9:48Like they have the tools you. Would think so, right?
- 9:51But the semi analysis report detailed in our sources
- 9:54forcefully dismantles that assumption.
- 9:56Simply owning hardware doesn't give you an advantage if your
- 9:59internal infrastructure strategy is a complete disaster.
- 10:02Really a disaster. Oh, a total disaster.
- 10:04If the infrastructure you build is fundamentally misaligned with
- 10:07the people actually writing the AI models, all that money just
- 10:10goes up in smoke. And the sources outline 3
- 10:12specific cases of this internal dysfunction, starting with the
- 10:15Revos acquisition. OK this one is wild.
- 10:18I read this part. So Meta spent $2.5 billion to
- 10:24acquire a chip startup called Revos.
- 10:27And Revos is built around RISCV, which is an open source hardware
- 10:32instruction set architecture. Think of RISCV as a freely
- 10:36available blueprint for building a computer chip, unlike the
- 10:39proprietary blueprints controlled by companies like ARM
- 10:42or Intel. But Meadow bought them primarily
- 10:45to acquire what's called Customer owned tooling.
- 10:48Yeah. And customer on tooling is
- 10:49essentially the logistical and software capability to manage
- 10:53chip manufacturing yourself directly with the fabrication
- 10:56plan. So you don't have to rely on the
- 10:58middleman. But the absurdity here is the
- 11:00price tag. The 2.5 billion, yes.
- 11:02Building that specific tooling capability from scratch
- 11:06internally at Meta would have cost roughly $100 million and
- 11:09taken about a year. Instead, they spent 2.5 billion.
- 11:12They overpaid by $2.4 billion just for a logistical
- 11:16capability. Which is insane, but you might
- 11:18think, well at least they got the startups cutting edge chip
- 11:20designs and they're brilliant engineering team right?
- 11:23No, Meta immediately cancelled the actual chip project Revos
- 11:28was working on. Revos had insisted on an all or
- 11:31nothing sale, meaning Meta had to buy every single employee
- 11:34even though they only wanted a tiny fraction of the logistics
- 11:37team. The sources use a brilliant
- 11:39analogy here. Oh.
- 11:40The F1 analogy. I'd love this.
- 11:42One, it's so good, they said. It's like buying a championship
- 11:45Formula One team. But when they arrive at your
- 11:47headquarters, you put the drivers in the accounting
- 11:49department, you put the aerodynamics, engineers and
- 11:51marketing, and you stick the pit crew in HR.
- 11:54And then management stands around screaming, asking why the
- 11:58company isn't winning car races. It was an immediate
- 12:00organizational rejection of a transplant.
- 12:03Other departments inside Meta just started cannibalizing the
- 12:06startup, grabbing engineers to build their own internal
- 12:09fiefdoms. A huge chunk of those highly
- 12:11specialized engineers ended up leaving in frustration or just
- 12:14getting laid off. That's rough.
- 12:16And that is just case one. Case 2, The Grand Teton server
- 12:19shows an even deeper disconnect between the people buying the
- 12:22hardware and the people actually using it.
- 12:25OK, so the infrastructure team at Meta designed a highly
- 12:28customized server for the Open Compute Project codenamed Grand
- 12:31Teton, and they made this unilateral decision to add an
- 12:35entire extra switch tray to the server design.
- 12:38This included four Broadcom PCIe switches and 16 extra
- 12:43solid-state drives. Right, and a PCIe switch is
- 12:46basically a high speed data highway interchange inside the
- 12:49server, directing traffic between the processors and the
- 12:52storage. So why did they add it?
- 12:54Well, the infrastructure team's logic was at the AI model
- 12:57training teams. We're going to need massive
- 12:59amounts of local physical storage right there on the
- 13:02servers to save their model checkpoints because training in
- 13:05AI model takes months, so you have to constantly save your
- 13:08progress or checkpoint it so a power outage doesn't wipe out
- 13:11$1,000,000 of compute time. Which I mean, sounds totally
- 13:14logical on paper. Right on.
- 13:15Paper, sure, but they never actually asked the AI model team
- 13:19if they needed local storage. The infrastructure team just
- 13:22assumed they knew best, so when they rolled these hyper
- 13:26expensive custom Grand Teton servers into production, the
- 13:30model team barely even touched that extra storage.
- 13:33They preferred saving their checkpoints to a centralized
- 13:36network storage system instead. So the entire custom design was
- 13:40a colossal waste A. Huge waste.
- 13:43It made the server significantly more expensive, physically
- 13:45hotter, and more complex to maintain, all to support a
- 13:49feature of the end user explicitly did not want.
- 13:52The design was eventually just scrapped entirely.
- 13:54It's textbook lack of Co design. It's like they basically built a
- 13:57massive industrial walk in freezer for a tenant who eats
- 14:00out for every single meal. They were burning electricity to
- 14:03keep empty shelves cold. That's the perfect way to
- 14:06describe. It but as bad as that is case, 3
- 14:08is definitely the most damaging for Meta's AI ambitions hobbling
- 14:12their own Ferraris. Oh yeah, this one is painful.
- 14:15Meta has this obsessive culture of customizing hardware, but
- 14:19historically they customize it to favor their legacy business.
- 14:23You know, the recommendation systems that power the Facebook
- 14:25and Instagram feeds. And a recommendation system
- 14:28requires a massive amount of CPU power or central processing
- 14:31units. That's because the algorithm is
- 14:33constantly doing table look UPS, searching through massive
- 14:36databases of your past clicks and likes to figure out what
- 14:39real to show you next. But generative AI like ChatGPT
- 14:44or Llama operates on a totally different mechanical level.
- 14:47It requires massive GPU power, graphics processing units, and
- 14:51incredible memory bandwidth. A large language model isn't
- 14:54just looking up past data, it is running billions of complex
- 14:57parallel mathematical calculations simultaneously to
- 15:01predict the next logical word in a sentence.
- 15:03GPUs are designed for that massive parallel processing, but
- 15:06they need to talk to each other constantly at lightning speed.
- 15:09Exactly. But Meta's infrastructure
- 15:11managers, they forced a custom server design called Aerial for
- 15:14Nvidia's cutting edge GB 200 chips.
- 15:17And against all logic for generative AI, they insisted on
- 15:20a one to one ratio of CPUs to GPUs.
- 15:23They deliberately bottlenecked the system to serve their old AD
- 15:26targeting workloads, basically starving the GPU's of the data
- 15:29they need to train efficiently. And they did the exact same
- 15:32thing to AMD. They ordered AMD to take their
- 15:35absolute bleeding edge MI450 chip, which is an engineering
- 15:39marvel design purely for AI, and literally demanded that AMD cut
- 15:43its compute dies and high bandwidth memory layers in half.
- 15:46It's just absurd. High Camera with memory or HPM
- 15:49is the ultra fast memory stacked directly on top of the GPU that
- 15:52allows the chip to access the data it needs instantly.
- 15:55Cutting that in half is like taking a Ferrari engine and
- 15:57installing a fuel line from a lawnmower.
- 16:00It doesn't matter how fast the engine can go if it can't get
- 16:02any gas. The most fascinating part of
- 16:05this, though, is the internal reaction.
- 16:07Meta actually has a dedicated frontier model lab, the TBD lab.
- 16:12These are the people whose literal job is to build the next
- 16:15generation of AI. How did they react to being
- 16:18handed this deliberately downgraded hardware?
- 16:21They absolutely despised it. The TBD lab wasn't even
- 16:24consulted before these multibillion dollar hardware
- 16:27decisions were locked in by the infrastructure procurement team.
- 16:30The AI scientists are looking at this incredibly expensive,
- 16:33highly customized, yet severely underpowered infrastructure and
- 16:37they want absolutely no part of it.
- 16:39The internal warfare is so bad that the Semi analysis report
- 16:42actually publicly urged AMD to bypass Meta's infrastructure
- 16:45managers entirely and try to sell the proper full strength
- 16:48chips directly to the TBD lab. So the procurement department is
- 16:52actively fighting a war against the engineering department,
- 16:55armed with billions of dollars. Here's what I want to know Why
- 16:59are some of the smartest, most highly compensated engineers in
- 17:02the world making such terrible, disjointed hardware decisions?
- 17:07Well, the root Causeway isn't technical incompetence.
- 17:10These are smart people. The issue is a deeply toxic
- 17:13corporate culture at Meta AI, driven by their human resources
- 17:17structure, specifically their performance review cycle and
- 17:22their reliance on brutal stack ranking.
- 17:24Stack ranking. That's an HR philosophy where
- 17:26managers are forced to grade their employees on a bell curve,
- 17:29right? No matter how incredible the
- 17:31team is, you are mathematically required to label a certain
- 17:34percentage of them as bottom performers.
- 17:36Yes, and at Meta they are cutting the bottom 10 to 15% of
- 17:40staff every single review cycle. Every cycle.
- 17:43That's brutal. It is, and furthermore, those
- 17:46performance reviews happen every six months.
- 17:49Now think about the mechanical timeline of hardware
- 17:51development, designing a custom silicon chip, securing
- 17:55fabrication space at TSMC, and building a new data center
- 17:59architecture That takes three to four years of sustained, focused
- 18:03effort. You literally cannot drive long
- 18:05term infrastructure projects with a six month survival of the
- 18:08fittest performance cycle. It creates an environment where
- 18:11everyone is just optimizing for their next review.
- 18:14Exactly. The sources refer to this
- 18:16cultural phenomenon as window washing.
- 18:18Middle managers are terrified of falling into that bottom 10% and
- 18:22losing their jobs. So instead of undertaking quiet
- 18:25necessary three-year foundational work, they forced
- 18:28through highly visible, flashy short term projects that look
- 18:32fantastic on a quarterly presentation.
- 18:34They wash the windows of the house.
- 18:35Looks shiny right now. They secure their massive
- 18:39bonuses, they survive the HR call, and then they immediately
- 18:42abandoned the long term execution of that project to
- 18:44start the next shiny thing. Oh I see, so they get all the
- 18:47credit for launching the spaceship, but they suffer 0
- 18:49consequences when it blows up halfway to Mars because they've
- 18:52already transferred to a new department?
- 18:53Precisely. And because of this chaos, the
- 18:57supply chain teams have absolutely zero power.
- 19:00They are constantly jerking outside vendors around, ordering
- 19:03huge custom chip designs and then canceling the orders a few
- 19:07months later when a new manager takes over.
- 19:09Here's where it gets really interesting though.
- 19:11Nobody dares challenge the senior leadership on this
- 19:14structural flaw. The sources mentioned a detail
- 19:17about the day-to-day reality at Meta that I think perfectly
- 19:20encapsulates this dysfunction. Despite spending $145 billion on
- 19:25servers, many software engineers don't even have assigned desks.
- 19:29Well, the desk thing, yeah, that is wild.
- 19:32You have these highly paid elite AI engineers wandering around
- 19:36the office every morning like they are at a busy coffee shop
- 19:39looking for a place to sit when they finally find an empty desk.
- 19:43Sometimes the monitors have the wrong display cables or there
- 19:46are no monitors at all, and when they apply through the internal
- 19:49system to get a dedicated desk, they get denied by automated HR
- 19:53software even though there are empty desks visible all around
- 19:57them. It really is the ultimate
- 19:58microcosm of the macro problem if a trillion dollar company
- 20:02cannot successfully provision a simple desk and a working
- 20:05monitor for an engineer. How can they possibly be trusted
- 20:08to successfully provision $100 billion custom AI data center?
- 20:14You have a thoroughly disjointed organization making hardware
- 20:17decisions that will define the next decade of technology.
- 20:20Right, So what does this all mean for you listening right
- 20:22now? The core take away here is a
- 20:24massive reality check on the AI industry and the era of $100
- 20:28billion hyperscalers. Simply throwing mountains of
- 20:31cash at hardware is not enough. No, it's really not.
- 20:34In fact, if your corporate incentives are misaligned, that
- 20:37money is actively destructive. If an infrastructure team is
- 20:41building systems based on the legacy workloads the company
- 20:44needed five years ago, and they aren't communicating with the
- 20:46scientists writing the AI models of tomorrow, those billions of
- 20:50dollars just vanish into thin air.
- 20:52Yeah, and it proves that in the AI boon, the biggest buyer is
- 20:56absolutely not the smartest buyer.
- 20:59Understanding the cultural rot that can fester beneath a
- 21:02trillion dollar market valuation is critical for anyone trying to
- 21:06navigate or invest in this space.
- 21:08It is a stark reminder to look past the hype and focus on the
- 21:11actual mechanics of execution. Which brings us to a final
- 21:14thought to leave you with today. We've just detailed how the
- 21:17initial infrastructure of this entire global AI revolution was
- 21:21built on an unstable foundation of panic buying, misaligned
- 21:24corporate incentives, and internal political turf wars.
- 21:29But what if this massive failure by the tech giants opens a
- 21:32completely different? Door like a different paradigm
- 21:34for AI development. Exactly what it's The next truly
- 21:38massive leap in artificial intelligence doesn't come from a
- 21:41trillion dollar hyperscaler blindly throwing $100 billion at
- 21:46custom servers in some shiny mega campus because Big Tech's
- 21:50architecture is proving to be so bloated, inefficient, and
- 21:53economically toxic. The next breakthrough might
- 21:56actually come from a lean cash strapped startup operating out
- 21:59of a garage. That's a fascinating point.
- 22:01A team that doesn't have the luxury of wasting a billion
- 22:04dollars on the wrong PCI E switches and therefore is forced
- 22:07to innovate incredibly efficient hyper focused software to run on
- 22:12standard hardware. Necessity breeds invention, and
- 22:15right now the Giants have absolutely no necessity.
- 22:18It's entirely possible that limitless budgets are the exact
- 22:21thing killing innovation at the top.
- 22:23Well, that's all for today's deep dive.
- 22:25Keep questioning the blueprints, track the real market, and we'll
- 22:28see you next time.