Our big reports last week on Marvell Technology (MRVL: NASDAQ), Box (BOX: NYSE), Okta (OKTA: NASDAQ), Synopsys (SNPS: NASDAQ), Veeva Systems (VEEV: NASDAQ), Docusign (DOCU: NASDAQ) and Hewlett Packard Enterprise (HPE: NYSE) are available to clients, with video summaries on YouTube and, of course, our popular summaries of the summaries on Instagram and Tiktok. (A calendar of our upcoming video reports at the end of this note). 

As always, I’m focused on three things:

1) Technology-driven change;
2) the latest startup trends;
3) and the Golden Age of Fraud. 

In This Note:

• Oracle's AI Flex

• TikTok Tech Trouble & Trump's X Win

• HAWK Tua: No Happy Ending

• HP's Mystery Customer & Backlog Bust

• Upcoming Earnings In Focus

• My Plans This Week

Oracle's AI Flex

Oracle (ORCL: NYSE) is pivoting hard to the cloud, and specifically to AI infrastructure. Q1 fiscal 2025 results showed it's working. This week’s Q2 2025 results will show even more.

In Q1, total cloud revenue hit $5.6 billion, up 22% in constant currency. CEO Safra Catz told us to expect cloud eclipsing license support revenue in Q2 for the first time. Leading the charge: 46% growth in infrastructure cloud services to $2.2 billion. (Full disclosure: I’m long ORCL.)

Oracle's strategy? “Gigantic GPU farms,” purpose-built for AI, says founder and CTO Larry Ellison. An 800-megawatt data center will house “acres of NVIDIA (NVDA: NASDAQ) GPU clusters” to train massive neural nets. 

But Oracle's approach differs from rivals Amazon (AMZN: NASDAQ) AWS, Microsoft (MSFT: NASDAQ) Azure and Alphabet’s (GOOG: NASDAQ) Google Cloud. More, smaller data centers - not fewer, bigger ones. 162 facilities, identical except for scale. Modular, automated, efficient.

This distributed model targets data residency and latency needs while scaling leading AI tech. Notably, Oracle's public and private clouds mirror each other precisely in hardware and features.

Power is a challenge. Oracle's answer: small modular nuclear reactors (SMRs). They'll supply a 1 GW data center requiring 3 SMRs. Regulatory hurdles remain, but it may be the only way.

The tech is dizzying:

• 32,000-node NVIDIA GPU clusters with 800 Gbps RDMA 

• 131,072 NVIDIA Blackwell GPUs coming in 2025

• 300+ exabytes of filesystem, 100+ exabytes of flash  

Oracle has emerged as a cloud innovator and AI infrastructure contender. With distinctive architecture, Oracle-only technologies (Exadata, ZFS) and an aggressive SMR roadmap, it's in the game and playing to win.

The zettascale era is rapidly on the way. 

TikTok Tech Trouble & Trump's X Win

TikTok's meteoric rise stems from algorithmic superiority. But this tech could enable Chinese interference in US politics and society. I've seen its addictive personalization firsthand as a creator (@DrillDownPod, please subscribe!).

On Friday, a US appeals court upheld a bipartisan law banning TikTok over national security, forcing divestment from ByteDae. The court prioritized security over free speech.

A Jan. 19 deadline looms, but could be delayed by a Supreme Court appeal or action by President Biden or incoming President-elect Trump. Microsoft, Oracle and others once vied for TikTok. (Full disclosure: I’m an Oracle shareholder.)

This will certainly be a boon for other creator platforms, such as Alphabet’s YouTube Shorts, Meta’s (META: NASDAQ) Instagram’s Reels and to a lesser degree Snap (SNAP: NASDAQ) and Meta’s Facebook Stories.

But TikTok won’t be easily recreated. The technology stack is jaw dropping. Deep learning models power TikTok's recommendation engine. They analyze user behavior, preferences, and video content using NLP and computer vision. By combining user and content insights, the system predicts irresistible videos for each person.

TikTok's artificial intelligence goes beyond generic interest matching. It assesses pacing, sounds, hashtags and visuals that resonate with individual tastes. The models spot emerging trends to surface viral-worthy content.

Enabling such complex AI and big data processing at TikTok's immense scale requires a colossal and expensive technical infrastructure. The company relies on a global network of data centers, each housing hundreds of thousands of servers, to power its platform. Running this infrastructure likely costs billions annually in equipment, electricity, cooling, and top AI/engineering talent. On the video delivery front, TikTok employs advanced compression techniques like H.264 and H.265 encoding to substantially reduce video file sizes without degrading quality. When a user opens the app, the content delivery network (CDN) swiftly retrieves videos from servers closest to the user's location, enabling smooth playback within seconds, even in lower bandwidth environments.

TikTok also optimizes the user interface for simplicity and responsiveness. By leveraging frameworks like React for web and native languages like Swift and Kotlin for mobile, the app delivers a snappy, intuitive experience designed to minimize friction and keep users engaged.

While TikTok's financials are private, its Chinese parent ByteDance has acknowledged the enormous operating expenses of its data-hungry algorithms, like tens of billions a year.

But TikTok's hyper-targeting and China's data laws pose huge risks. In 2018, ByteDance gave China a stake in TikTok's Chinese sister app. And that’s worrisome, given China’s open hostility towards the United States and its stated goal to take over Taiwan. 

Social media algorithms have already demonstrated their electoral impact. During the 2016 U.S. election, Russian operatives reached 126 million Facebook users through strategic content amplification and misinformation that outpaced legitimate news sources. Facebook's 2018 algorithm change to prioritize "meaningful social interactions" further revealed how such tweaks could amplify polarizing content during critical voting periods. 

In TikTok's case, ByteDance's control enables even more sophisticated possibilities - from subtly suppressing news about Chinese military exercises to amplifying content questioning Taiwan's independence. During a potential Taiwan crisis, the algorithm could systematically downplay invasion news, spread confusion about international responses, or amplify anti-intervention voices, all while maintaining deniability through its complex recommendation system.

We have very real and very recent experience in America with a social media network tweaking its algorithms to favor certain voices in favor of one extreme election result: this year Elon Musk had X’s engineers alter algorithms so that his political posts received more views than all U.S. political campaign ads on X's disclosure dataset. 

In July alone, after Musk’s endorsement of Trump, his political posts had 17.1 billion views “over two times as many views as all the US ‘political campaigning ads’ that X has recorded in its political ads disclosure dataset in the same period” said The Center for Countering Digital Hate.

Did Musk's algorithm changes favoring Trump on X sway the race? You have to ask yourself. Does social media influence people? Does advertising work? How much would it take to make a difference? Trump's 1.62% win was the slimmest since 2000. 

This experience highlights social media's vulnerability to foreign meddling. Absent separation from ByteDance, TikTok remains a conduit for election interference and surveillance. Its innovative tech is hugely valuable, informative and fun, but the national security danger is undeniable. The US must act decisively to protect Americans' data and democracy from foreign influence. Securing TikTok's potent information and algorithms is critical.

HAWK Tua: No Happy Ending

Haliey Welch, the social media star who went viral for her “Hawk Tuah” video, is giving crypto investors something else to swallow - a crypto token that collapsed in what appears to be one of the largest memecoin schemes to date. And though she denies selling any of her tokens, someone has minted millions through a Cayman Islands trust that Welch apparently had a hand in creating.

“HAWK”, launched December 4 on the Solana blockchain, saw its market capitalization surge to $490 million before plummeting 94% to $25 million in just hours. The token's 15% transaction fee generated approximately $2.1 million for the Southern Skies Foundation, a Cayman Islands entity, based on the $14 million in its first day of trading volume.

The token's distribution raised red flags from the start. Of the 10 billion HAWK tokens, supposedly 30% went to a “treasury reserve,” 21% to a “community fund,” 20% to “Haliey’s fans”, 17% to "strategic partners”, 10% to Haliey herself (supposedly locked for 12 months), and 2% for public allocation.

But blockchain analysis reveals 17% of the tokens were “pre-sold” to some entity. And a bubblemaps blockchain analysis shows that at least 79% of HAWK's supply was controlled by insider wallets and “snipers” — entities that acquire large positions at launch only to dump their holdings. One such sniper wallet purchased 17.5% of the memecoin's supply for about $993,000, later selling 135.8 million tokens for a $1.3 million profit within two hours.

On Wall Street, we’d call that a “pump-and-dump” but in cryptoland it’s a “rugpull” — promoters pulling the rug out from investors. 

Behind the scenes, HAWK demonstrated technical sophistication now common in crypto schemes. The token's creation followed a precise sequence: First, developers deployed a Solana Program Library (SPL) token contract, minting 10 billion HAWK tokens at a cost of approximately 5 SOL ($500). Two billion of these tokens were distributed by Welch and her 18-person team between Nov. 26 and Dec. 2, given to fans, her social media followers and even podcast guest Mark Cuban, giving the appearance of widespread ownership and interest in the token. 

When trading opened on decentralized exchanges like Raydium, the initial liquidity pool was seeded with just 50 SOL paired against 10 billion HAWK tokens, creating an artificial price baseline. As trading volume grew, the contract's automated 15% fee accumulated in designated wallets. At precisely 3:47 PM EST, these fees were converted to SOL through multiple simultaneous trades across different liquidity pools, using Solana's near-instant transaction finality to complete the conversion before arbitrage traders could react.

Welch denied selling any tokens, stating her allocation remains locked. However, blockchain records show massive selling from wallets associated with the “community fund” and “strategic partners” allocations.

While her podcast and YouTube exploits should have prepared followers for disappointment, HAWK’s plunge left a particularly bitter taste for Hawk Tuah’s crypto-believing fans.

HP's Mystery Customer & Backlog Bust

Blowout quarter. Puzzling backlog bust. Hewlett Packard Enterprise (HPE: NYSE) posted record Q4 fiscal 2024 revenue of $8.5 billion. Infrastructure cloud services soared 49%. Compute revenue hit an all-time high. Then came the bombshell.

HPE took a $700 million order "de-book" from an unnamed customer. Net AI orders plunged from $1.2 billion to $500 million. Executives cited vague "risk" but shared few specifics. Who was this mystery buyer?

The 2023 10-K may held few clues: "The Company has one customer which represented 11% of the Company's total net revenue in fiscal 2023, primarily within the Intelligent Edge and Compute segments." It’s unlikely that a customer that spent $788 M with HPE became so undependable the next year, but who knows. 

The murky disclosure jars with the AI spending boom. NVIDIA (NVDA: Nasdaq), whose GPUs power HPE's AI systems, is struggling to meet demand. Cloud giants are racing to build mammoth language models. Why retreat?

Theories abound. Perhaps the cancelled order involved NVIDIA's unproven Blackwell "AI superchip", due in 2025. Over one-third of HPE's $3.5 billion AI backlog is Blackwell-based.

Or just a matter of timing? CFO Marie Myers emphasized AI's “lumpy” nature amid “a pretty competitive environment.” With NVIDIA and AMD (AMD: NASDAQ) launching rival accelerators, the mystery customer may be waiting to see how things unfold.

Big picture, HPE remains bullish on AI infrastructure demand. The company is constructing massive data centers, including an 800-megawatt facility that will house “acres of NVIDIA GPU clusters” to train expansive neural networks. HPE also has plans for data centers exceeding 1 gigawatt in power consumption.

And importantly, in the period between the quarter’s end and the quarterly conference call, HPE assured us that bookings were reflated back to 3.5 billion. Nothing to see here kids, move along.

But in the race to an AI future, even fast-growing AI players will hit bumps. Safety belts advised.

Upcoming Earnings In Focus

My Plans This Week

I'll be in San Francisco all week and quite available (slight chance I’ll be with a client in the midwest late this week or next.)

I hope these thoughts are helpful to you. I’d love to discuss them further and, as always, comments, questions and ideas are appreciated.

/