<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Yizhi Liu on Medium]]></title>
        <description><![CDATA[Stories by Yizhi Liu on Medium]]></description>
        <link>https://medium.com/@yzhliu?source=rss-6a992ef60c85------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*cM9uC9_40X12BV5eAcrmzA.jpeg</url>
            <title>Stories by Yizhi Liu on Medium</title>
            <link>https://medium.com/@yzhliu?source=rss-6a992ef60c85------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 16:39:34 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@yzhliu/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[What P/E Are We Really Paying for Tech Stocks?]]></title>
            <link>https://medium.com/@yzhliu/what-p-e-are-we-really-paying-for-tech-stocks-3b647c0f0bb9?source=rss-6a992ef60c85------2</link>
            <guid isPermaLink="false">https://medium.com/p/3b647c0f0bb9</guid>
            <category><![CDATA[business]]></category>
            <category><![CDATA[palantir]]></category>
            <category><![CDATA[nvidia]]></category>
            <category><![CDATA[valuation]]></category>
            <category><![CDATA[investing]]></category>
            <dc:creator><![CDATA[Yizhi Liu]]></dc:creator>
            <pubDate>Mon, 16 Feb 2026 01:48:05 GMT</pubDate>
            <atom:updated>2026-03-03T05:15:31.148Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>managements sometimes assert that their company’s stock-based compensation shouldn’t be counted as an expense. (What else could it be — a gift from shareholders?)<em><br>— Warren Buffett, 2018 Annual Letter to Shareholders</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bm8umVhdKq82YxvesfYZag.png" /></figure><p>After my previous article on Palantir’s business model, a reader asked why Michael Burry believes Palantir has never actually made money in its history. Let’s talk about that today.</p><p>The table below comes from Palantir’s FY2024 annual report. Focus on the Common Stock/Shares column. From 2021 to 2024, the number of common shares outstanding was 2.027 billion, 2.099 billion, 2.2 billion, and 2.338 billion, growing by 3.55%, 4.81%, and 6.27% respectively. In other words, over three years, investors’ ownership was diluted by roughly 15%.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*P9HsFA6-XplHCuih" /></figure><p>Where did all these newly issued shares go? The same table tells us plainly: option exercises and RSUs. U.S. tech companies compensate employees with stock — in the early stages through options that grant the right to purchase shares at a set strike price, and after going public, directly through restricted stock units (RSUs). The distinction is mainly driven by tax considerations, which I won’t get into here.</p><p>These equity awards, as reflected in the table, are not paid out of treasury stock. They are brand-new shares — a direct dilution of existing investors’ ownership. The question is: how should we account for this impact?</p><p>Intuitively, these equity payouts are part of employee compensation and should be treated as an operating cost. U.S. GAAP indeed requires exactly that. The table includes an item called Stock-based Compensation (SBC), which came to $565 million, $476 million, and $693 million in 2021, 2022, and 2023 respectively. This expense is allocated across Cost of Sales, Sales and Marketing, R&amp;D, and G&amp;A on the income statement. So the Net Income you see has already been reduced by SBC.</p><p>Since SBC is not a cash expense, it gets added back in the cash flow statement, as shown below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*I0_OhOk5P4dNG596" /></figure><p>Company valuations should be based on actual cash inflows and outflows, not accounting profits — that’s common sense and consensus. In 2024, Palantir’s full-year net income was $468 million. Add back the $692 million in SBC, and operating cash flow breaks through $1 billion. That’s how Wall Street analysts do the math.</p><p>But is that math correct?</p><p>Of course not. Remember what we discussed at the beginning — to fund SBC, Palantir shareholders saw their equity diluted by 15% over three years, roughly 4.6% per year. An expense that causes a real decline in equity value cannot simply be “added back.”</p><p>We need to push further and ask: is that $692 million even the true cost of SBC?</p><p>When a company grants equity to employees, RSUs are typically valued at the closing price on the grant date, while option valuations are far more complex. But regardless of the method, the figure booked is not the same as the actual cost borne by shareholders.</p><p>To arrive at the true cost of SBC, let’s go back to that original equity table. In 2024, Palantir conducted a buyback: 2,123,000 shares for a total of $64,196,000, at an average cost of $30.24 per share. Meanwhile, the total number of new shares issued for SBC in 2024 was 140 million. If the company wanted to buy back enough shares to offset all newly issued equity, the total cost would be a staggering $4.257 billion. Keep in mind, 2024 adjusted profit was only $468 million + $691 million (non-cash SBC) = $1.16 billion — roughly one-quarter of the cash needed for that buyback. Looking back one year, even if we assume Palantir could have repurchased at an average price of $10 per share in 2023, the buyback cost would still reach $1 billion, far exceeding that year’s adjusted income of $690 million. This enormous gap is exactly what Michael Burry means when he says “they don’t make anything.”</p><p>By now, many of you might be thinking: is this really how you’re supposed to calculate it? If you still have doubts, let’s look at a real-world example of a company that actually uses buybacks to offset equity dilution: NVIDIA (the following calculations are taken entirely from Michael Burry’s analysis).</p><p>Since 2018, NVIDIA has spent $91 billion buying back its own stock. Yet by 2025, NVIDIA’s total share count hasn’t decreased — it has even ticked up slightly. Let’s assume the buybacks roughly offset shareholder dilution. Subtract the hard cash spent on buybacks from reported profits, and NVIDIA’s actual cumulative profit since 2018 is $135 billion, not the $205 billion reported in its financials. Factor in deferred taxes associated with SBC, and that number drops further to $114 billion — approximately 56% of reported profits. In other words, this company’s real P/E is 60x, not the 30x your stock app displays.</p><p>So much of this is just common sense: if you pay the bulk of salaries in stock, does the expense magically vanish just because shares can be issued “out of thin air”? That defies intuition, defies common sense, and is simply not the right thing to do.</p><p>I often criticize the systemic preferential treatment of passive income. In truth, distributing equity to the employees actually doing the work is both fair and necessary for technological innovation. But an honest company shouldn’t be evasive about this, and a responsible company should know that when you earn $1.1 billion in a year, you cannot hand out $4.2 billion worth of stock.</p><p>Berkshire Hathaway recently opened a position in Google. Many people said they “couldn’t understand it” or that “it doesn’t look like a Buffett decision.” While the decision may indeed not have come from Buffett himself, if you take Google’s financials and run the same analysis, you’ll understand why Google, like Apple, is considered a value benchmark among tech stocks. Value investing emphasizes nothing more than the most fundamental principles of private ownership, chief among them:</p><p><em>Corporate ownership has a price tag, and to acquire ownership, one should pay the corresponding price.</em></p><p>This might sound like a truism at first glance, yet many brilliant minds either fail to understand it — or understand it but cannot bring themselves to act on it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3b647c0f0bb9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Economics of Chinese LLMs: Analyzing the Financials of Zhipu and MiniMax]]></title>
            <link>https://medium.com/@yzhliu/the-economics-of-chinese-llms-analyzing-the-financials-of-zhipu-and-minimax-2bfebefdfbfe?source=rss-6a992ef60c85------2</link>
            <guid isPermaLink="false">https://medium.com/p/2bfebefdfbfe</guid>
            <category><![CDATA[investing]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ipo]]></category>
            <category><![CDATA[buisness]]></category>
            <dc:creator><![CDATA[Yizhi Liu]]></dc:creator>
            <pubDate>Fri, 16 Jan 2026 07:56:07 GMT</pubDate>
            <atom:updated>2026-01-16T07:56:07.891Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Economics of Chinese LLMs: the Financials of Zhipu and MiniMax</h3><p>Two Chinese LLM leaders, Zhipu AI and MiniMax, just went public on the Hong Kong Stock Exchange — Zhipu on January 8, MiniMax the day after. You’ve probably already seen plenty of coverage rehashing what’s in their prospectuses, so I’m not going to waste your time with that. What I want to do instead is actually crack open their financial data and look at how these companies make and spend money. Because honestly, all this hand-wringing about whether AI is in a “bubble” means nothing if we’re not looking at actual numbers.</p><h4>Inference Costs Remain High</h4><p>Let’s start with gross margins, because they tell you a lot about what’s actually happening under the hood. MiniMax was bleeding money on every sale in 2023. The gross margin clawed back to 12.2% in 2024 and hit 23.3% in 2025, which is progress, but still pretty thin. Zhipu’s numbers look way better at first glance: they’ve been sitting comfortably at 50–60% for the past three years. But here’s the thing: that’s mostly because nearly 70% of their business comes from ToB and ToG contracts that require on-premise deployment, meaning their customers are footing the bill for inference costs, not Zhipu. When you look at their cloud-based deployment, where Zhipu actually has to pay for the compute, the gross margin in the first half of 2025 was negative.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-k8G6pP0rp9a4T9D1pWkAw.png" /><figcaption>Zhipu AI’s gross profit and gross profit margin breakdown</figcaption></figure><p>Given the current state of AI infrastructure (chips, servers, data centers, and electricity), <strong>the low-margin reality for the C-end (consumer market) is unlikely to change in the short to medium term. Costs aren’t coming down, and expansion is constantly hitting computing power bottlenecks. The AI ToC business is very difficult.</strong></p><p>While shifting to the B-end and G-end offers higher margins than ToC, a 60% gross margin is hardly “excellent” — it’s merely “average.” Industry benchmarks like Palantir exceed 80%, and Salesforce sits at 70%. A margin of 50%–60% puts these startups on par with players like IBM or C3.ai.</p><p>Furthermore, as business scales, <strong>Accounts Receivable turnover has lengthened significantly:</strong> from 23 days in 2022 to 112 days in the first half of 2025. By Silicon Valley standards, that doesn’t sound great. But let’s be real about the market they’re operating in. When your main customers are Chinese SOEs and government agencies, 112 days is actually pretty decent. It’s frustrating, but it’s the reality of doing business in this ecosystem. The real question is whether they can keep this from getting worse as they grow, or if we’re going to see these payment cycles stretch out to six months, nine months, or longer. That’s something to keep an eye on.</p><p>The product is delivered, but the money hasn’t arrived — a headache the ToC world doesn’t have.</p><h4>Runaway R&amp;D Spending</h4><p>Anyone reading these prospectuses would find it hard to ignore the staggering R&amp;D expenditures — the figures are so large they practically jump off the page.</p><ul><li>MiniMax: R&amp;D spending in 2023 and 2024 reached $70 million and $189 million, respectively (compared to revenues of just $3.5 million and $30 million).</li><li>Zhipu: R&amp;D spending for 2022, 2023, and 2024 (converted at 7:1 RMB/USD) was $12 million, $75 million, and $314 million, respectively (against revenues of $8.2 million, $17.8 million, and $44.6 million).</li></ul><p>So where does all this money go? Computing power for model training. Zhipu’s data shows that computing power accounts for 70.7% of R&amp;D, with salaries coming in second at 14.8%.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ozrURQIl7AYHJMJgNslxGw.png" /><figcaption>Zhipu AI’s R&amp;D expense breakdown</figcaption></figure><p>Developing foundation models is often seen as the core competitiveness of an AI company. However, as model capabilities converge and it becomes harder to maintain a lead, differentiated competition disappears.</p><p><strong>The Best Case:</strong> After several rounds of brutal competition, most players exit, leaving two or three winners who can iterate on general capabilities while waiting for computing costs to fall.</p><p><strong>The Worst Case:</strong> The “Scaling Law” hits a wall, and as costs drop, you don’t just face your current competitors — you face everyone. More startups pile in, and open-source models flood the market. At that point, <strong>running a foundation model company becomes like running a gravel pit.</strong> Your rocks aren’t special. Your excavator isn’t special. You’re just another player in a commodity business where the only competitive lever left is who can burn money the slowest. <strong>And if you know anything about commodity businesses in truly competitive markets, you know that’s where profits go to die.</strong></p><p>And if it turns out that actually useful applications still need tons of industry-specific data to work properly, these companies will eventually have to answer a tough question: How do you justify hundreds of millions of dollars in annual training costs for a foundation model?</p><h4>Cash-Burning Advertisements</h4><p>Even if we ignore R&amp;D, net profits are being dragged down by marketing costs.</p><p>MiniMax’s marketing expenses in 2023 and 2024 were $23 million and $87 million — 6.5 times and 2.9 times its annual revenue, respectively. <strong>This is still the old “burning cash for traffic” playbook from the mobile internet era. The problem is: has this purchased traffic created a strong enough network effect? </strong>In the first nine months of 2025, after significantly cutting ad spend, MiniMax’s average MAU dropped by 34%, and new user acquisitions plunged by 50%. These results are hardly satisfactory.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ENI91KTidtHRaDeaUwcYgA.png" /><figcaption>MiniMax’s MAU details</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3YegJh9LjVQHj4tDvg8xsg.png" /><figcaption>MiniMax’s new users details</figcaption></figure><p>The situation at Zhipu isn’t much better; marketing expenses are roughly equal to revenue. <strong>This reflects the general state of the Chinese SaaS industry: gross margins around 50%, coupled with high marketing costs and long AR turnover. Part of this is the business environment. But from another perspective, AI commercialization has yet to find a disruptive path in the short term.</strong> Chinese SaaS has always been a tough slog; now it’s just been rebranded as MaaS (Model as a Service). New name, same old pitfalls, plus new problems brought by computing costs.</p><h4>Does Open Source Have a Business Model?</h4><p>A quick word on open source. Many people assume by inertia that open-source models can be commercialized, citing successes like RedHat Linux, Docker, or Apache Spark. However, there is a fundamental difference: <strong>Open-source LLMs lack an ecosystem.</strong></p><p>Traditional open-source software works because of community dynamics. Sure, most users just download and use it for free. But there’s always a subset of developers who give back: they file issues, fix bugs, build new features. More importantly, they become evangelists who bring in more users, which speeds up development. It’s a virtuous cycle — the more people use it, the better the software gets. And the company leading the project? They make money by selling services, support, and enterprise features around that free core.</p><p>The key to real open source is transparency — you can see how it was built. But these so-called “open-source” models? There’s no training data. No training code. They just dump the model weights and call it a day. It’s like if I wrote a C++ program, compiled it into a binary, and then announced I’m “open-sourcing” the .exe file. <strong>That’s not open source. That’s just free software.</strong></p><p>Similarly, when others take your open-source model and fine-tune it, they don’t return their data or training code to your ecosystem. Compared to traditional software, the ecosystem for open-source models is so much weaker. When you factor in the costs, a viable business model doesn’t exist. <strong>In the end, open-sourcing a model is just a PR stunt — essentially turning massive R&amp;D costs into marketing expenses.</strong></p><p>Though, there is still value in technical accumulation.</p><h4>Beyond the Numbers</h4><p>I often wonder: how should we view a technology bubble?</p><p>In the U.S., long-term capital gains are taxed at a maximum of 20%, while personal income tax reaches 37%. Corporate taxes have dropped from 52.8% to 21%, yet they still fund a massive amount of social welfare, healthcare, and pension costs. We live in a world where the “idle rich” often have more tax advantages than the hardworking average person. <strong>In a system that institutionally encourages passive income, how do we address the question of fairness?</strong></p><p>In a way, technology does create a kind of equality: it democratizes access. Whether you’re rich or poor, you can use Google. You can learn to code online for free. The infrastructure doesn’t care about your bank account. And over the past 50 years, tech has been the most powerful force for scrambling the wealth deck. It keeps creating new ladders for people to climb and kicking out the rentiers who thought they could just coast on their existing assets forever.</p><p><strong>A tech bubble is simply an intense way of accelerating this process of technological democratization.</strong> In a recent interview, Jeff Bezos stated that the current AI boom should be seen as an “industrial bubble” rather than a purely “financial bubble.” A financial bubble, like the 2008 banking crisis, is “all bad” for society. But an industrial bubble, even when it bursts, leaves behind a valuable legacy.</p><p><strong>The destruction of value brought by a burst bubble is never the destruction of technical value. It only destroys greed and laziness.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2bfebefdfbfe" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Scaling Trap: Why Internet Growth Model is Failing in the AI Era]]></title>
            <link>https://medium.com/@yzhliu/the-scaling-trap-why-internet-growth-model-is-failing-in-the-ai-era-e39f4a3e0b5b?source=rss-6a992ef60c85------2</link>
            <guid isPermaLink="false">https://medium.com/p/e39f4a3e0b5b</guid>
            <category><![CDATA[business]]></category>
            <category><![CDATA[investing]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Yizhi Liu]]></dc:creator>
            <pubDate>Sun, 21 Dec 2025 23:47:17 GMT</pubDate>
            <atom:updated>2025-12-21T23:47:17.244Z</atom:updated>
            <content:encoded><![CDATA[<p>In November 2022, ChatGPT was launched. In just two short months, it became the fastest-growing application in human history to reach 100 million users. To achieve the same milestone, TikTok took nine months, and Instagram took two and a half years.</p><p>Many viewed this as proof that AI applications were taking off. People began duplicating the “Internet-era playbook” for AI: rapid products launch, aggressive user acquisition, and continuously iterating to <strong>scale up</strong>. A flood of consumer AI apps sprouted up — particularly chat-based apps like Character.ai, which captured massive traffic in a very short time.</p><p>However, it did not last long. Only a year and a half after a $150 million funding round and entering the $1 billion “unicorn” club, Character.ai’s management and core model team departed for Google. This de facto acquisition clearly signaled the market’s stance at the time: the application itself wasn’t the prize; the technology was.</p><p>Similar stories played out among Chinese AI startups. After several rounds of aggressive user acquisition and product scaling, star AI companies scaled back their product layouts. The typical strategies favored by Chinese internet giants suddenly proved ineffective in the AI era.</p><p>So, why?</p><p>To answer this, we must look past the phenomena to re-examine the essence of some past success stories.</p><p>When people think of the “Internet growth model,” the first thing that comes to mind is often low pricing. In 2017, when I traveled from San Francisco Airport to San Jose, there were still a few yellow cabs waiting for passengers. I remember that 40-mile trip cost about $140 by taxi, while Uber was less than half that price. While I didn’t witness Uber’s earliest phases, I vividly remember the 2014 subsidy war between Didi and Kuaidi, two Rideshare Apps operated in China, which made rides essentially free at the time. This capital-driven scaling soon pushed traditional taxi companies out of the market.</p><p>Looking at Internet companies that disrupted traditional industries, the core element of success was the so-called “<strong>network effect</strong>.” Take rideshare as an example: the more passengers use Uber, the more drivers are attracted to the platform. This makes getting a ride easier, which in turn leads more people to choose Uber, forming a closed loop. In other words, <strong>as the scale expands, the product automatically becomes better</strong>. You don’t even need to add new features; you just need to focus on scaling. For nearly 20 years, WeChat’s primary interface is still just a chat window, it’s an ultimate expression of the network effect.</p><p>In the later stages of the Internet era,<strong> the efficiency of scaling</strong> reached its zenith. First, Internet reached the most remote regions through fiber optics and base stations. Then came the rise of cloud computing, reducing server expansion cycles from weeks or months to mere minutes. The growth of open-source collaboration platforms like GitHub allowed companies to deploy high-quality infrastructure software at almost zero cost. In 2009, during my internship at a game company, I observed how servers took weeks to deliver to production to scale just for one game. Today, scaling up is simply a single kubectl command. This was unimaginable 20 years ago.</p><p>Meanwhile, the price of unit computing has continued to drop. I asked AI to summarize computing prices over the past 20 years, while not accurate, it offers a glimpse into the macro trend:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WisTdP9NQpzBpHbbGzYl1A.png" /></figure><p>Bolstered by both infrastructure and pricing, Internet companies were able to scale freely. If scaling itself improves product quality, and scaling is incredibly cheap, then the success of these products is no surprise. They only need a bit of capital to provide the initial momentum for the scale flywheel; after that, they just leave it to what code does best: self-replication.</p><p>Once we understand this, it’s easy to see why the same logic is hard to apply to AI.</p><p><strong>First, the network effect is weakened.</strong> Many chat applications are even anti- network effects initially: if the AI is the only conversation partner, users are actually isolated from one another. An expanding user scale doesn’t yield a “1+1 &gt; 2” result. While the app may understand users better as data accumulates, this is a very long-term process. Furthermore, it is constantly hampered by a more severe issue:</p><p><strong>Scaling has reverted to a timeline of weeks or months, a throwback to the way things were twenty years ago.</strong> Renting cutting-edge GPU clusters from cloud providers usually requires a minimum six-month contract; even then, in many regions, delivery times can drag on for months. For those building their own data centers, a two to three months delay in GPU servers isn’t uncommon. The expansion of hyperscalers is further limited by data center construction and power supply, often requiring planning years in advance. These factors significantly increase both the time and financial costs.</p><p>As these advantages diminish, the product flywheel driven by scale simply stops spinning.</p><p>But, regardless of external factors, what matters most is the value of the product itself. In 2000, startups like PayPal and Amazon faced similar scaling difficulties due to immature infrastructure, yet they successfully maintained high user retention while scaling. I don’t see such level of value creation in many applications today (and of course, in 2000 too).</p><p>In the long run, the golden age of AI applications will demand smaller, more energy-efficient computing units. If the power of an H100 could be packed into a smartphone today, it would be impossible for AI applications not to take off. <strong>Whether Moore’s Law can hold up, and where the physical limits of chip manufacturing actually lie, are perhaps more vital problems to solve than launching GPUs into space.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*oyVkZwsJH1IYTaGS.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e39f4a3e0b5b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Palantir Playbook]]></title>
            <link>https://medium.com/@yzhliu/the-palantir-playbook-fdb545b53757?source=rss-6a992ef60c85------2</link>
            <guid isPermaLink="false">https://medium.com/p/fdb545b53757</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[palantir]]></category>
            <category><![CDATA[stock-market]]></category>
            <category><![CDATA[business]]></category>
            <dc:creator><![CDATA[Yizhi Liu]]></dc:creator>
            <pubDate>Fri, 12 Dec 2025 04:54:39 GMT</pubDate>
            <atom:updated>2025-12-13T01:57:59.783Z</atom:updated>
            <content:encoded><![CDATA[<p>Palantir stock has delivered over 140% year-to-date, and a staggering 20x since its 2020 listing, corresponding to an annualized growth of over 80%. Researching Palantir seems to have become a mainstream obsession in investment circles. Especially at a time when people are struggling to find AI value in the market, who can say no to a high-growth AI company that is continuously generating positive cash flow?</p><p>Let’s set aside the rationality of the share price for a moment and look at a comparison of some fundamental data:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xokq_vzfkxVymLWQPlEQtg.png" /></figure><p>Whether in terms of profit margins, capital efficiency, or growth, Palantir rivals or surpasses top-tier SaaS providers. Meanwhile, traditional defense contractors, dragging their bloated bureaucratic systems, have chosen to “lie flat.” As for C3.ai, whose business is quite similar, the company has been open for 16 years yet its operating cash flow remains negative — proving that frequently changing your name doesn’t necessarily make you rich.</p><p>To understand a business, one must grasp the major aspects while letting go of the minor ones. Many factors affect a company’s short-term performance, but in the long run, the core elements number no more than two or three. As an outsider, all one can do is piece together the company’s most fundamental business logic from specific cases.</p><p>The collaboration with Airbus is a case Palantir takes great delight in discussing. It was mentioned in four different places in their first annual report for fiscal year 2020 and is frequently cited by executives and sales staff in public settings. This partnership began in 2015. According to Palantir, they imported Airbus’s internal factory schedules, staff shifts, parts, delivery, and defect data into the Foundry platform. Through data analysis and monitoring, they helped the Airbus team accelerate A350 deliveries by 33%.</p><p><strong>What exactly is this Foundry platform?</strong></p><p>The Palantir website features a set of cool animated graphics, paired with even more confusing codenames and jargon, seemingly trying very hard to explain it to you. Yet, the more you read, the less you understand what’s inside or how it operates. It brings to mind a classic line from the movie <em>The Big Short</em>: “Wall Street loves to use confusing terms to make you think only they can do what they do. Or, even better, for you to just leave them the fuck alone.”</p><p>It’s actually quite simple: <strong>this is a data analytics system deeply customized for each client.</strong></p><p>Every enterprise is special, with its own unique internal culture, processes, and efficiency bottlenecks. Furthermore, many enterprises haven’t even completed the most basic digitization; vast amounts of information are stored as Word documents, scanned PDFs, or even paper printouts. “Garbage in, Garbage out” is every AI practitioner’s nightmare. Even after painstakingly cleaning the data and structuring it into a form machines can understand, designing models and processes to derive reasonable insights requires a profound understanding of the client’s business and production workflows. Expecting to solve the efficiency needs of all enterprise-side users in a standardized, platform-based way is unrealistic, even with today’s AI capabilities.</p><p>Such deep customization leads to several important, positive consequences.</p><p>First, it ensures relatively significant output. To be fair, given the level of digitization in many traditional enterprises (including government departments), as long as you can navigate key political nodes and are willing to put in the “dirty work,” a 30% efficiency improvement isn’t too difficult — especially since the metrics used to measure results are themselves negotiable.</p><p>More importantly, spending massive time and efforts on data integration brings high customer stickiness. The migration cost for private data is extremely high; as more data gets integrated into this internal platform, it becomes very difficult for the enterprise to break away. This is also why Palantir is willing to do data processing and structuring for free in the early stages of a project, because these early investments yield continuous cash flows for years to come.</p><p>The story of Palantir and Airbus doesn’t end there. In 2016, Palantir extended this internal platform to more departments. By 2017, Palantir and Airbus further expanded the project, developing software called <strong>Skywise</strong>, and then sold this software to airlines operating Airbus planes and to Airbus suppliers. Every time Airbus sells a plane, it bundles Skywise to the airline. Skywise is not cheap, and it charges based on the number of seats, in the manner of traditional SaaS.</p><p>If you observe Palantir’s most successful cases, including the UK’s NHS and Japan’s SOMPO Holdings (an insurance company with a large elderly care business), they all follow this pattern: start with an industrial giant, deeply customize their internal solutions, and then use this client as a sales channel to push standardized, platform-ready functions out to the client’s clients.</p><p>There are many outsourcing companies that customize software for clients, but few can sell custom software at SaaS prices, and even fewer can turn clients into sales distribution channels. Alex Karp said, <em>“Most enterprise software is bought by people who don’t use it, for people who hate it. Our software is used by the people who love it.”</em> It’s self-promotion. But it is undeniable: <strong>this is the power of deep customization based on an understanding of the client’s business.</strong></p><p>Some might ask, how do you scale custom software? The answer is simple: scalability doesn’t have to be the kind as, for example, social media apps. As long as gross margins reach a certain level, this is more than fine. Software already scales way better than other heavy-asset, labor-intensive businesses. <strong>On the contrary, pursuing so-called “platformization” too early and at too high a level is one major pitfall for enterprise business.</strong></p><p>In a recent interview, Michael Burry said Palantir is essentially no different from IBM’s business. I agree. But compared to who is increasingly “lying flat” and losing its technological edge, Palantir is obviously doing better.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wPwz7wqQKVxkf1cUiWnc6A.png" /></figure><p>As for Michael’s comment that “Historically, they (Palantir) don’t make anything.” I may save that for another article later. Speaking of which, I can’t help noticing how these star AI companies react to even a tiny bit of skepticism nowadays — cursing on earning calls or frantically publishing posts during trading hours, as if they’ve suffered some grave injustice. Their poise is really lacking.</p><p>End of the day, no matter how excellent a company is, it must eventually return to a reasonable value. Carrying a P/E of four or five hundred, it’s better to be a little humble.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fdb545b53757" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>