Previously we published a post about the practice of Capitalizing R&D software development costs and outlining the pros and cons of doing so to help guide the decisions of what, when and how much to capitalize versus expense.

As outlined in our original post, unfortunately we see many companies "over-capitalize" R&D in order to make profitability look better in the short to medium term. Capitalizing R&D results initially in a more favorable P&L (income statement) since a smaller amount of R&D expense shows up in the current quarter (for public companies) and more of the original software development cost is deferred to future quarters. The total value of the work capitalized to create the feature/functionality then sits on the Balance Sheet and is depreciated over each period of its designated lifespan. This deferral of expense really only makes financial sense when the product being developed is truly new/innovative (and meets the GAAP criteria) and will therefore have a useful life that is aligned with the depreciation period roughly. Stated another way, if the software development cost to create the new feature/functionality meets the GAAP criteria to capitalize, it should also provide competitive advantage in most cases for a similar period of time. For example, if the investment is depreciated over 3 years, it should usually have competitive business value for roughly/at least 3 years in the marketplace. Unfortunately for many going concern businesses that heavily capitalize most new feature development, the effort being capitalized is a stretch at best to meet the GAAP criteria for capitalization treatment. In addition, if the level of ongoing R&D expense is relatively steady (the Engineering team size is stable), then eventually the depreciation in each period is the same as it would have been without capitalizing since the level of depreciation eventually reflects the ongoing investment aligned to a stable team size. Once depreciation drag is close or the same as it would have been if the company had expensed most R&D on an ongoing basis, the company is simply adding an unnecessary expense of administration as well as audit risk, based on the choice to capitalize the software development cost rather than simply expense it in the current reporting period/quarter.

While the financial realities outlined above are unchanged, as time has passed, regulatory, political and social climates relative to these behemoth companies have evolved and shifted. Therefore it’s worthwhile to re-analyze the current state of capitalizing versus expensing R&D costs overall in an ever-changing environment. In addition, with the acceleration in Generative AI (GenAI) advancements and investment levels, we will also explore the accounting treatment of exploding AI related product development costs. Specifically, given the escalating size of Capital budgets by all the major US tech players, we wanted to understand if they are applying a different philosophy to their capitalization practices based on the level of investment, experimentation and relative to the idea of potential longer enduring value.

These updates will be comprised of a two-part series of posts. Part I will be focused on investments by a subset of the established Tech ‘behemoths’ for both core ongoing products as well as generative AI at a high level (1. below), while Part II will focus more on Generative AI specifically in terms of fundamental model development, training and operation as well as NVIDIA as the leading supplier for microchips in the emerging AI space (2. and 3. below).

As part of this update on software development accounting practices, we will paint a full picture by examining the following three questions.

  1. How are the large incumbent tech companies currently treating their software development costs regarding expense versus capitalization?

    • How are the established tech players treating their R&D costs overall?
    • Has anything changed materially since we last analyzed this question?
    • What is their outlook regarding Capex? And relative GenAI capex? Which portions are they capitalizing?
    • What traction in the marketplace are these companies achieving thus far as it relates to GenAI?
    • What are these companies strategies for chip development?

    We base our analysis on the recent financial statements and earnings announcements from Meta, Google, Microsoft and Amazon (based on the last 2 quarters of the 2024 calendar year). We include Amazon and Microsoft due to their size and public cloud presence even though much of their LLM current strategy is through partnerships, with Anthropic (Amazon) and OpenAI (MSFT). AI costs are not consistently or explicitly broken out in many cases, so we rely heavily on Notes to Financial Statements and Management comments as part of Earnings Call Transcripts from the two most recent quarters.

    In part II we will also examine studies to evaluate the investments directly in fundamental model development.

  2. How are AI related costs specifically being treated? Since some of the leading GenAI companies developing fundamental models are private, financial data is limited to what is available from recent funding rounds or through partnership data with Amazon and MSFT (and potentially others such as Oracle).

    • How are model training costs being treated?
    • How are costs roughly split between model development and model training?
    • Similarly what is the split of costs between compute and human development costs?
    • What may change going forward with costs to build and maintain foundational models and why?
    • Based on the limited data available for OpenAI’s 2024 funding round, can we draw any conclusions on differences between their cost treatment and the established players’ treatment?

  3. What are the investment levels as well as accounting treatment of R&D costs at NVIDIA? It’s almost impossible to have a conversation about AI costs without mentioning the costs of GPU chips as well as the growing NVIDIA platform to run AI applications on. Therefore, it’s useful to understand:

    • Nvidia’s R&D cost profile and growth over time (both revenue and cost).
    • Accounting treatment of NVIDIA’s R&D investments
    • Do they capitalize more as one might expect some of their investments, given hardware (chips) and lack of real competition, would be expected to have more enduring value than that of the software/services focused tech giants?
    • Do they separate and/or treat hardware related development costs differently than software?
    • Although it can’t be perfectly separated, what is their rough split between software and hardware development costs? And how may this shift over time?

Part I: Update on Accounting Treatment of R&D investments at the largest Tech Companies

META key findings

  • Meta continues to expense most of its “baseline R&D spend” relative to human effort in the current period/quarter. This is evident based on the total R&D expense and associated Notes to their Financial Statements.
    • This philosophy is further supported by their relatively low Intangible assets, which is the typical Balance Sheet line item for capitalized R&D expenditures being gradually amortized.
  • Meta is significantly increasing its R&D spend, despite recent restructuring (layoffs), largely driven by:
    • Wage inflation
    • Higher infrastructure costs
  • GenAI related infrastructure investments are driving the increase in Capex levels, both absolute and relative to past periods.
  • The significant capital investments, primarily in infrastructure, to develop, train and run LLM models will add to ongoing depreciation that is realized quarterly, creating a longer tail or ongoing increase in quarterly R&D cost over time.
  • Meta breaks their GenAI investments roughly as follows:
    • Core AI – categorized as AI to run the current business (ads), which ‘already exists’ as a category of product capabilities for Meta
    • Gen AI – future opportunities across their business with the ‘next gen AI model capabilities/power’
      • Meta is NOT expecting investments in the "GenAI category” to yield positive returns in the near term. This is a typical innovation philosophy to help stave off the Innovators Dilemma, especially for larger companies, i.e. if Meta put stringent requirements for near term return on GenAI investment, they would never make the investment and would lose the longer term arms race for AI dominance.
      • GenAI is however expected to start to ‘meld’ into core AI in 2025/26, suggesting that Meta expects returns to start to align to investment in the Horizon 2 to early Horizon 3 timeframe.
      • Meta has been consistently increasing its Capex spending guidance while reporting each quarter’s results and attributing the driver of this capex increase is allocated to GenAI investments.
      • GenAI investments currently have a larger capex component compared to ongoing software development due to the upfront cost to build very large compute clusters powered with specialized chips.
  • Chip development - in April 2024 Meta announced the Meta Training and Inference Accelerator (MTIA) chip designed to optimize AI model training and inference tasks. Meta is also deploying Artemis, which is a custom-designed AI processor. They have no plans to sell either of these chips externally, therefore the current effort is to reduce reliance on Nvidia, reduce cost and optimize for their use cases.

Google (Alphabet) key findings

  • Google is still expensing most of its R&D spend related to workforce costs to build software. This is evident based on the total R&D expense in their P&L/income statement. This conclusion is also backed up by the Notes to their financial statements and the comments in the most recent 2 quarters' earnings calls.
  • Although Google does not use the category “Intangible Assets” on their Balance Sheet which is a common place to categorize capitalized software assets, they do likely capitalize most R&D related infrastructure investments under the balance sheet category “Property and Equipment.” Some infrastructure costs may also be included in “Operating lease assets” line item. Lastly there may be a small amount of other R&D people expense under the “Other Assets” categories, but this is a less common place to put these costs and they state clearly that of R&D costs, the main capex cost contribution, is based on infrastructure costs.
  • While Alphabet does not SPECIFICALLY say they do not capitalize a significant amount of R&D (as Meta does state explicitly), Alphabet does say that their capex is mainly tech infrastructure based on their most recent quarterly reports. Therefore, we conclude that that depreciation component of R&D expense on the Income Statement that is amortized quarterly – is primarily the drawdown in infrastructure book value (as opposed to software amortization).
  • Google’s organic search business is obviously both a risk and an opportunity for GenAI, but currently more skewed toward risk given the perception that the initial Search Overviews launch was rushed.
  • Of the major incumbent players, Google does have arguably the longest and deepest foundation of AI to build upon with its many years of investment in DeepMind and other research efforts.
  • Google is also giving guidance to higher Capex levels going forward, but to a lesser extent than the other players (some of the other incumbents are increasing Capex by a larger percentage).
  • Chip Development - Similar to Amazon, Google has a heavy focus on incremental AI investment relative to its GCP business, both in terms of NVIDIA chips as well as their own homegrown Tensor Processing Unit (TPU) chips (latest version is Trillium, TPU v6e as of Oct 2024). Tech infrastructure investment for major GenAI datacenter buildouts is roughly 60% chips/40% rest of datacenter/networking. Given the compute intensity and therefore cost of running GenAI services and applications, the ability to improve performance more economically is critical to both their core search and public cloud businesses. Having their own chips that are truly performant relative to the market leader, is obviously critical to their ongoing competitiveness in terms of cost and capabilities.

MSFT key findings

  • MSFT appears to capitalize relatively more of its ongoing R&D spend, compared to Meta/Google, based on the comments to their financial statements stating that they expense costs up to the point of technical feasibility and subsequently capitalize costs until the product is generally available to customers.
  • Capitalized development depreciation is included in the Cost of Revenue line item on their Income Statement.
  • Their R&D expense includes ongoing development and only depreciation of purchased software.
  • MSFT is also increasing their capex guidance, heavily driven by the GenAI arms race.
    • MSFT expects that their capex is relatively higher right now then it will be ongoing (as a percentage of revenue) with the acceleration of GenAI innovation and investment. They talk regularly in their earnings call about initial significant investment followed by realizing ‘ongoing inference value.’
    • MSFT shows more “clear-cut current revenue” for their GenAI investments based on their success in selling/adding on Copilot, compared to the other incumbents. Copilot as a product is more “naturally visible and transparent” than GenAI products of their competitors since customers buy and experience Copilot directly as a consumer. In contrast, Meta’s AI products are one step removed from customers' direct use and experience. Google’s AI products are currently ‘free’ as part of search or part of the GCP business and therefore less transparent in terms of revenue/value. Similarly, Amazon’s AI efforts are also removed and/or free as part of the retail experience or less transparent as part of AWS. Therefore, while Copilot has measurable sales, it is not possible to directly compare MSFT’s market success to AI market success of their competitors (and of course MSFT is deploying AI in their cloud business as well as across all of their businesses in addition to Copilot)
    • The CEO and CFO tout growth of copilot being the fastest MSFT business to reach $10 billion.
    • The management team also suggests that smaller businesses are buying more O365 partially attributed to the (additional) utility of Copilot for small business use cases/automation.
    • MSFT has been relatively successful out of the gate with Copilot sales, but editorially Copilot has recently received quite a bit of negative feedback that its value/usefulness is poor compared to the current license cost (expect there may be a ‘pull back’ in enterprise’s willingness to pay for Copilot, keep the same number of seats, etc.)
  • Chip development – MSFT is investing in its Maia 100 and Cobalt 100 chip development to support AI model development and inference at a performant cost-effective scale. They are also heavily collaborating with OpenAI but this does not appear to include chip development as of yet.

AMAZON key findings

  • Amazon is not specific regarding the split between capitalization of the R&D effort itself in addition to the infrastructure investment (which is usually capitalized). Based on their Income Statement notes, their general philosophy is to expense payroll related costs for ongoing feature development in the current period but capitalize infrastructure investments. Both the ongoing expense and the depreciation expense are included in their “Technology & Infrastructure” line item in their P&L/Income Statement.
  • However, Amazon does state that they sometimes capitalize development on larger initiatives once they have reached the stage of commercial viability. For example, in a recent earnings announcement they use their global broadband service investment as an example of a product that will be partially capitalized once it reaches commercial viability.
  • In the most recent quarter, Amazon also indicated that they have reduced depreciation by extending the useful life of some infrastructure investments. They do not specify which infrastructure this change applies to. One hypothesis is that extended infrastructure asset life could apply to emerging GenAI investments, related to being Anthropic’s primary cloud provider or their own GenAI innovation/model development, but this accounting change could apply to any or all infrastructure.
  • Similar to the other large tech players, Amazon is giving guidance to expect (relative) increasing levels of capex in coming quarters (near to medium term).
  • Amazon has recently invested an additional $4 billion in Anthropic, bringing its total investment level to $8 billion.
  • Chip development – Amazon is investing in developing its own chips which currently include Trainium for model development and training and Inferentia, for inference tasks once a model is running (utilizing AWS’s Annapurna Labs). They have announced publicly that Anthropic will use Amazon chips as part of their growing partnership. This is of note compared to MSFT that has NOT made any announcements about OpenAI’s use of MSFT developed chips.

OVERALL KEY TAKE-AWAYS

  • The big tech incumbents have not changed their policy to, for the most part, expense costs related to human effort on software development in the current period or quarter.
  • META and Google especially indicate that they expense most R&D ongoing software development costs. This is not surprising given their more ‘pure play’ software products, roots and heritage.
  • Amazon also expenses ongoing R&D human costs based on their most recent reports but also capitalizes larger research efforts which is appropriate and expected.
  • MSFT appears to capitalize more effort related to ‘ongoing feature development’ based on their Financial Statement Notes and their Income Statement but still expenses most ongoing R&D investment cost as it relates to human effort.
  • The level of capitalization is not possible to quantify perfectly for any of these 4 companies based on their financial statements.
  • ALL of these companies ARE significantly increasing absolute and relative capital spend. The primary driver of this increase is cited as GenAI investment, concentrated in infrastructure – no surprise there!
  • GenAI investment is VERY capital intensive given the need for specialized more expensive chips that are in limited supply and the need to build out more data centers to support the accelerating demand (and/or support their own competitiveness assuming actual demand does not match the level of buildout yet, in order to achieve profitability).
  • Based on comments to Financial Statements and Earnings call transcripts, the Capex is primarily attributed to infrastructure as opposed to capitalizing the cost of human effort to build and run models and new GenAI applications. This again supports the conclusion that the incumbent players are treating the development costs as an expense in the same way they do in their core business. We will evaluate the cost of actual large fundamental model development separately in Part II of this post.
  • This ‘capital glut’ resulting primarily from AI infrastructure investment, will result in relatively higher depreciation levels, which impacts ongoing profitability until demand is at a level where revenue exceeds ongoing costs including infrastructure depreciation.
  • Most companies expect the higher level of relative capital investment to reduce to “more normal ongoing levels of capital investment” in line with revenue growth, standard expense and investment ratios etc. The timing for a ‘return to normal’ is very unclear though given the arms race to lead or be left behind in achieving market success with GenAI capabilities. Investors naturally have some anxiety about this situation as is evidenced by questions during Earnings calls and much talk about “have we reached the Trough of Disillusionment for GenAI yet”?
  • The coming increase in ‘depreciation drag’ on ongoing P&L, as a result of higher than normal capex investments, could be partially offset by extending useful life for this growing area but this is unlikely given the current fast-paced cycle of innovation

  • NO companies are achieving profitability with ‘true GenAI’ investments yet but each have their own stated path(s) to profitability.
    • Meta with their ‘melding’ of core and GenAI capabilities over time, likely heavily related to Adtech
    • Google with their use of Gemini in organic search, their growing GCP cloud segment and other products.
    • MSFT is perhaps the most successful of the group thus far on the path to GenAI business success - with a ‘real product’ (for both Consumers and Businesses) to point to in Copilot. However as previously stated, overall satisfaction is mixed at best regarding Copilot living up to its hype and productivity promise relative to its current cost, suggesting there may be some pullback or at least slow down in expansion. MSFT also has a relatively wider product portfolio that will benefit from AI as it relates to segments such as Azure and gaming. In addition, MSFT’s partnership with OpenAI is still early and therefore its business success is unclear.
    • Amazon’s GenAI opportunities are heavily focused on AWS and their philosophy to be flexible in hosting (most) major models. They are also making significant investments in their own Trainium chips to manage cost/performance tradeoffs. Amazon is of course also developing GenAI applications to improve the retail experience but again, given their AWS focus and profitability, near term GenAI business opportunities for both Amazon and their customers are AWS centric.

In summary, the tech behemoths have not changed their approach to managing R&D investments overall and expensing most of the software development effort on a quarter-by-quarter basis, BUT they are making significant incremental capital allocations primarily in infrastructure to run GenAI at scale. The profitability of these investments is still generally beyond the current time horizon but some players are showing more progress than others. Join us in Part II of this exploration when dive into model development costs and accounting! For more insights and strategies on navigating complex technology investments, consider attending our Free CTO Accelerator Webinars, designed to equip technology leaders with the tools they need to succeed.